Network Traffic Classification: From Theory to Practice

Comments

Transcription

Network Traffic Classification: From Theory to Practice
Network Traffic Classification:
From Theory to Practice
Valentı́n Carela-Español
Advisor: Dr. Pere Barlet-Ros
Co-Advisor: Prof. Josep Solé-Pareta
Ph.D. in Computer Science
Universitat Politècnica de Catalunya BarcelonaTech
Department d’Arquitectura de Computadors
Barcelona, October 2014
Curso académico:
Acta de calificación de tesis doctoral
Nombre y apellidos
Programa de doctorado
Unidad estructural responsable del programa
Resolución del Tribunal
Reunido el Tribunal designado a tal efecto, el doctorando / la doctoranda expone el tema de la su tesis doctoral
titulada ____________________________________________________________________________________
__________________________________________________________________________________________.
Acabada la lectura y después de dar respuesta a las cuestiones formuladas por los miembros titulares del
tribunal, éste otorga la calificación:
NO APTO
APROBADO
(Nombre, apellidos y firma)
NOTABLE
SOBRESALIENTE
(Nombre, apellidos y firma)
Presidente/a
Secretario/a
(Nombre, apellidos y firma)
(Nombre, apellidos y firma)
(Nombre, apellidos y firma)
Vocal
Vocal
Vocal
______________________, _______ de __________________ de _______________
El resultado del escrutinio de los votos emitidos por los miembros titulares del tribunal, efectuado por la Escuela
de Doctorado, a instancia de la Comisión de Doctorado de la UPC, otorga la MENCIÓN CUM LAUDE:
SÍ
(Nombre, apellidos y firma)
NO
(Nombre, apellidos y firma)
Presidente de la Comisión Permanente de la Escuela de Secretario de la Comisión Permanente de la Escuela de
Doctorado
Doctorado
Barcelona a _______ de ____________________ de __________
A mis padres y hermanos,
Abstract
Since its inception until today, the Internet has been in constant transformation.
The analysis and monitoring of data networks try to shed some light on this huge
black box of interconnected computers. In particular, the classification of the
network traffic has become crucial for understanding the Internet. During the last
years, the research community has proposed many solutions to accurately identify
and classify the network traffic. However, the continuous evolution of Internet
applications and their techniques to avoid detection make their identification a
very challenging task, which is far from being completely solved.
This thesis addresses the network traffic classification problem from a more
practical point of view, filling the gap between the real-world requirements from the
network industry, and the research carried out in the network traffic classification
field. In this work, we identify several problems that hinder the introduction of
new proposals for traffic classification in production networks. First, proposed
techniques usually do not meet the special requirements of production networks,
with massive amounts of traffic and limited resources to process it, which make
them very difficult to deploy. Second, classification techniques need to be updated
regularly in order to maintain its accuracy. The cost of this maintenance process
is usually unfeasible in networks with hundreds or thousands of routers. Last but
not least, there is a clear lack of systematic methods to validate and compare the
state-of-the-art techniques, mainly because most proposals base their results on
private datasets labeled with techniques of unknown reliability.
The first block of this thesis aims to facilitate the deployment of existing techniques in production networks. To achieve this goal, we study the viability of using
NetFlow as input in our classification technique, a monitoring protocol already implemented in most routers and switches. Since the application of packet sampling
has become almost mandatory in large networks, we also study its impact on the
classification and propose a method to improve the accuracy in this scenario. Our
results show that it is possible to achieve high accuracy with both sampled and
vii
viii
unsampled NetFlow data, despite the limited information provided by NetFlow.
Once the classification solution is deployed it is important to maintain its accuracy over time. Current network traffic classification techniques have to be
regularly updated to adapt them to traffic changes produced by the continuous
evolution of the Internet traffic. The second block of this thesis focuses on this
issue with the goal of automatically maintaining the classification solution without
human intervention. Using the knowledge of the first block, we propose a classification solution that combines several techniques only using Sampled NetFlow as
input for the classification. Then, we show that classification models suffer from
temporal and spatial obsolescence and, therefore, we design an autonomic retraining system that is able to automatically update the models and keep the classifier
accurate along time.
Going one step further, we introduce next the use of stream-based Machine
Learning (ML) techniques for network traffic classification. In particular, we propose a classification solution based on Hoeffding Adaptive Trees. Apart from the
features of stream-based techniques (i.e., process an instance at a time and inspect it only once, with a predefined amount of memory and a bounded amount
of time), our technique is able to automatically adapt to the changes in the traffic
by using only NetFlow data as input for the classification. In order to extract
sound conclusions we evaluate our technique using a 13 years long dataset from a
transatlantic link in Japan.
The third block of this thesis aims to be a first step towards the impartial validation and comparison of state-of-the-art classification techniques. The wide range
of techniques, datasets, and ground-truth generators make the comparison and validation of different traffic classifiers a very difficult task. To achieve this goal, we
evaluate the reliability of different Deep Packet Inspection-based techniques (DPI)
commonly used in the literature for ground-truth generation. Although according
to conventional wisdom they are, in theory, one of the most accurate techniques,
the results we obtain show that some well-known DPI-based techniques present
several limitations that make them not recommendable as a ground-truth generator in their current state.
In addition, we publish some of the datasets used in our evaluations to address
the lack of publicly available datasets and make the comparison and validation of
existing techniques easier. In particular, the dataset we publish in this last third
block is the first reliable labeled dataset with full packet payload available for the
research community. To the best of our knowledge, this is the only dataset available
that makes possible the comparison and validation of ML and DPI techniques.
Resumen
Desde sus orı́genes hasta la actualidad, Internet ha estado en constante evolución.
El análisis y la monitorización de las redes tratan de arrojar luz sobre la caja
negra de ordenadores interconectados que es Internet. En especial, la clasificación
de tráfico de red se ha vuelto crucial para su comprensión. Durante los últimos
años, la comunidad investigadora ha propuesto múltiples soluciones para identificar
y clasificar con precisión el tráfico de red. Sin embargo, la continua evolución de
las aplicaciones de Internet y sus técnicas para evitar ser detectadas, hacen que
su identificación sea un tarea muy complicada, y lejos de estar completamente
resuelta.
Esta tesis aborda el problema de la clasificación de tráfico de red desde un
punto de vista más práctico, tratando de hacer confluir la investigación llevada
a cabo y las necesidades de los entornos reales. En este trabajo identificamos
diferentes problemas que entorpecen la introducción de las nuevas propuestas para
clasificar el tráfico en redes troncales. En primer lugar, las técnicas propuestas ven
dificultado su despliegue dado que habitualmente no reúnen los requisitos necesarios para operar en redes troncales, con enormes volúmenes de tráfico y limitados
recursos para procesar. En segundo lugar, las técnicas de clasificación necesitan
ser actualizadas regularmente con el objetivo de mantener su precisión. El coste
de este mantenimiento es normalmente inviable en redes con cientos o miles de
enrutadores. Por último, pero no por ello menos importante, existe una gran falta
de métodos para validar y comparar las técnicas propuestas en la literatura. Esto
es debido, principalmente, a que la mayorı́a de propuestas basan sus resultados en
conjuntos de datos privados etiquetados con técnicas de desconocida fiabilidad.
El primer bloque de esta tesis pretende facilitar el despliegue de las técnicas
de clasificación en redes troncales. Para ello, estudiamos la viabilidad de usar
NetFlow, un protocolo de monitorización implementado en la mayorı́a de enrutadores y conmutadores del mercado, como entrada de nuestra técnica de clasificación. Además, dado que la aplicación de muestreo de paquetes es una práctica
ix
x
muy extendida en las redes troncales, estudiamos su impacto en la clasificación y
proponemos un método para mejorar su precisión en ese escenario. Los resultados muestran que, a pesar de la limitada información que nos proporciona NetFlow, es posible conseguir una alta precisión tanto con datos muestreados como
no muestreados.
Una vez desplegado el sistema de clasificación, el siguiente objetivo es mantener su precisión a lo largo del tiempo. Las soluciones actuales requieren actualizaciones periódicas para adaptarse a los cambios en el tráfico. El segundo bloque
de esta tesis se centra en este problema, persiguiendo automatizar el proceso de
mantenimiento y hacerlo sin intervención humana. Siguiendo la lı́nea del primer
bloque, proponemos un sistema de clasificación que combina varias técnicas que
usan únicamente NetFlow como entrada para la clasificación. A partir de este sistema mostramos que los modelos de clasificación sufren de obsolescencia temporal
y espacial y, para ello, diseñamos e implementamos un sistema de reentrenamiento
autónomo capaz de actualizar automáticamente los modelos y mantener la clasificación precisa a lo largo del tiempo.
Yendo un paso más allá, introducimos el uso de técnicas de Aprendizaje Automático (ML, por sus siglas en inglés) orientadas a flujos de datos para la clasificación de tráfico de red. En particular, proponemos una solución basada en
Hoeffding Adaptive Trees. Además de las caracterı́sticas propias de las técnicas
orientadas a flujos de datos (i.e., inspección única de cada instancia, con una cantidad de memoria predefinida y en un tiempo limitado), nuestra técnica es capaz de
adaptarse automáticamente a los cambios en el tráfico usando únicamente datos
NetFlow como entrada para la clasificación. Por último, para extraer conclusiones
sólidas, evaluamos nuestra técnica usando una traza que comprende 13 años de
tráfico de un enlace transatlántico en Japón.
El tercer bloque pretende ser un primer paso hacia la validación y comparación
imparcial de las propuestas del estado del arte. El amplio rango de técnicas, conjuntos de datos y generadores de verdad terreno, hacen la comparación y validación
de los diferentes clasificadores de tráfico una tarea muy complicada. Con ese fin,
evaluamos la fiabilidad de diferentes técnicas basadas en Inspección Profunda de
Paquetes (DPI, por sus siglas en inglés) habitualmente usadas en la literatura para
la generación de la verdad terreno. Aunque estas técnicas son, en teorı́a, unas de
las más precisas, los resultados que obtenemos muestran que algunas técnicas DPI
presentan graves errores que desaconsejan su uso en su estado actual.
Además, para abordar la falta de conjuntos de datos disponibles para la comunidad, publicamos algunos de los usados en nuestras evaluaciones para facilitar la
comparación y validación de las técnicas existentes. En particular, el conjunto de
xi
datos publicado en el tercer bloque es el primer conjunto de datos etiquetado de
manera fiable y con el contenido completo que está disponible para la comunidad
investigadora. Hasta donde sabemos, este es el primer conjunto de datos público
que permite la validación y comparación de técnicas ML y DPI.
xii
Resum
Des de els seus orı́gens fins l’actualitat, Internet ha estat en constant evolució.
L’anàlisi i el monitoratge de les xarxes tracten d’aclarir la caixa negra d’ordinadors
interconnectats que és Internet. Especialment, la classificació de tràfic de xarxa ha
esdevingut crucial per la seva comprensió. Durant el últims anys, la comunitat investigadora ha proposat múltiples solucions per identificar i classificar amb precisió
aquest tràfic. No obstant això, la contı́nua evolució de les aplicacions d’Internet i
les seves tècniques per evitar ser detectades, fan que la seva identificació sigui una
tasca molt complexa, i encara no completament resolta.
Aquesta tesi aborda el problema de la classificació del tràfic de xarxa des d’un
punt de vista més pràctic, tractant de fer confluir la recerca duta a terme i les
necessitats dels entorns reals. En aquest treball identifiquem diferents problemes
que dificulten la introducció de les noves propostes per classificar el tràfic en xarxes
troncals. En primer lloc, les tècniques propostes veuen dificultat el seu desplegament atès que habitualment no reuneixen els requisits necessaris per operar en
xarxes troncals, amb enormes volums de tràfic i limitats recursos per processar.
En segon lloc, les tècniques de classificació necessiten ser actualitzades regularment
amb l’objectiu de mantenir la seva precisió. El cost d’aquest manteniment és normalment inviable en xarxes amb centenars o milers de commutadors. Per últim,
però no per això menys important, existeix una gran falta de mètodes per validar
i comparar les tècniques propostes en la literatura. Això és degut, principalment,
al fet que la majoria de propostes basen els seus resultats en conjunts de dades
privades etiquetats amb tècniques de desconeguda fiabilitat.
El primer bloc d’aquesta tesi pretén facilitar el desplegament de les tècniques
de classificació en xarxes troncals. Per això estudiem la viabilitat d’utilitzar NetFlow, un protocol de monitorització implementat en la majoria de commutadors
del mercat, com entrada de la nostra tècnica de classificació. A més, donat que
l’aplicació del mostreig de paquets és una pràctica molt estesa en les xarxes troncals, estudiem el seu impacte en la classificació i proposem un mètode per millorar
xiii
xiv
la seva precisió en aquest escenari. Els resultats mostren que, malgrat la limitada
informació que ens proporciona NetFlow, és possible aconseguir una precisió prou
alta tant amb dades mostrejades com no mostrejades.
Una vegada desplegat el sistema de classificació, el següent objectiu és mantenir
la seva precisió al llarg del temps. Les solucions actuals requereixen actualitzacions
periòdiques per adaptar-se als canvis en el tràfic. El segon bloc d’aquesta tesi se
centra en aquest problema, perseguint automatitzar el procés de manteniment
i fer-ho sense intervenció humana. Seguint la lı́nia del primer bloc, proposem
un sistema de classificació que combina diverses tècniques que usen únicament
NetFlow com a entrada per a la classificació. A partir d’aquest sistema mostrem
que els models de classificació pateixen d’obsolescència temporal i espacial i, per
a això, dissenyem i implementem un sistema de reentrenament autònom capaç
d’actualitzar automàticament els models i mantenir la classificació precisa al llarg
del temps.
Anant un pas més enllà, introduı̈m l’ús de tècniques d’Aprenentatge Automàtic
(ML, per les seves sigles en anglès) orientades a fluxos de dades per a la classificació
de tràfic de xarxa. En particular, proposem una solució basada en Hoeffding
Adaptive Trees. A més de les caracterı́stiques pròpies de les tècniques orientades
a fluxos de dades (i.e., inspecció única de cada instancia, amb una quantitat de
memòria predefinida i en un temps limitat), la nostra tècnica es capaç d’adaptar-se
automàticament als canvis en el tràfic utilitzant únicament dades NetFlow com a
entrada per a la classificació. Per últim, per extreure conclusions sòlides, avaluem
la nostra tècnica utilitzant una traça que conté 13 anys del tràfic d’un enllaç
transatlàntic del Japó.
El tercer bloc pretén ser un primer pas cap a la validació i comparació imparcial de les propostes de l’estat de l’art. L’ampli rang de tècniques, conjunts
de dades i generadors de ground-truth, fan de la comparació i validació dels diferents classificadors de tràfic una tasca molt complicada. Amb aquest fi, avaluem
la fiabilitat de diferents tècniques basades en la Inspecció Detallada de Paquets
(DPI, per les seves sigles en anglès) habitualment utilitzats en la literatura per
la generació del ground-truth. Encara que aquestes tècniques són, en teoria, unes
de les més precises, els resultats que obtenim mostren que algunes tècniques DPI
presenten errors prou greus que desaconsellen el seu ús en l’estat actual.
A més, per abordar la falta de conjunts de dades disponibles per a la comunitat, publiquem alguns dels utilitzats en les nostres avaluacions per facilitar la
comparació i validació de les tècniques existents. En particular, el conjunt de dades
publicat en el tercer bloc d’aquesta tesi és el primer conjunt de dades etiquetat de
manera fiable i amb el contingut complet que està disponible per a la comunitat
xv
investigadora. Pel que sabem, aquest és el primer conjunt de dades disponible
públicament que permet la validació i comparació de tècniques ML i DPI.
xvi
Acknowledgments
Looking back I would not have been able to complete this long journey without
the help and support of countless people.
First of all I would like to express my gratitude to my advisor Dr. Pere Barlet
Ros, not only for being my mentor but for his expertise, support and patience
throughout my PhD studies. I must also show my sincere gratitude to my coadvisor Prof. Josep Solé Pareta for his support and encouragement in endless
aspects of this work. I honestly appreciate you both believed in me at the beginning
of this rewarding experience.
I am indebted to Dr. Kensuke Fukuda for selecting me for an internship at
the National Institute of Informatics (NII) in Tokyo. His expertise and the indepth discussions with Dr. Johan Mazel and Dr. Romain Fontugne, significantly
contributed to the knowledge used in creating this thesis.
I must acknowledge Dr. Alberto Dainotti and Prof. Antonio Pescapè for hosting me at Napoli University Federica II and Dr. Marc Solé Simó, our collaboration
in k-dimensional trees was very fruitful and a great learning experience.
Thank you to Dr. Albert Bifet for his guidance in the stream-based machine
learning field and for his extremely dedication to solving my doubts.
I would also like to thank Dr. Albert Cabellos-Aparicio for his constructive
discussions and support during the study of sampling in network traffic classification.
A special thanks to Dr. Tomasz Bujlow for his devotion in our fruitful collaboration in the study of the reliability of Deep Packet Inspection techniques.
Thanks to all the monitoring group: Dr. Ignasi Paredes, Dr. Josep Sanjuas,
Jakub Mikians, Georgios Dimopoulos, Ismael Castell Uroz and Oriol Mula for
assisting me during this long experience. Special acknowledgment to Oriol Mula
for our rewarding collaboration in the development of the Autonomic Retraining
System.
Thanks to Juan Molina for the exhaustive and precise work done in his master
xvii
xviii
thesis, being his advisor has been a gratifying experience.
I am very grateful to the external reviewers of this thesis Dr. Christian Kreibich
and Dr. Jens Myrup Pedersen as well as the members of the pre-defense committee
Prof. Jordi Domingo, Dr. Davide Careglio and Dr. Jordi Nin for their useful
comments that considerably help to improve this thesis.
I would like to acknowledge CESCA, UPCnet and MAWI for allowing the
collection of the network traffic traces used in this thesis. I thank ipoque for
kindly providing access to PACE and supporting the thesis of Juan Molina. I
also thank the European COST Action IC0703 and the NII for giving me the
opportunity of doing an internship in Napoli and Tokyo, respectively.
This thesis was mainly funded by a UPC research grant as well as by the Spanish Ministry of Education (contract TEC2011-27474) and the Catalan Government
(contract 2009SGR-1140).
I must thank all my friends for making easier this long journey: office mates,
lunchmates, gaijins, erasmitos, fibers, lechuguinos and, of course, deges. Their
friendship has been more important to this thesis than one might imagine.
Finally, I want to express my most sincere gratitude to my family for being
always there. I owe all of this to them, in particular to my little brother David,
he makes my days!
Thank you all. You have my most heartfelt gratitude.
Barcelona, October 2014
Valentı́n Carela-Español
Contents
List of Figures
xxv
List of Tables
xxviii
List of Acronyms
xxix
1 Introduction
1.1 Motivations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Contributions and Impact . . . . . . . . . . . . . . . . . . . . . . .
1.3 Thesis Organization . . . . . . . . . . . . . . . . . . . . . . . . . . .
2 Background
2.1 Network Traffic Classification Approaches
2.1.1 Port-based Approach . . . . . . . .
2.1.2 Payload-based Approach . . . . . .
2.1.3 Flow features-based Approach . . .
2.1.4 Host-behavior-based Approach . . .
2.1.5 IP-based Approach . . . . . . . . .
2.1.6 Comparing Approaches . . . . . . .
2.2 Network Traffic Classification Evaluation .
2.2.1 Performance Metrics . . . . . . . .
2.2.2 Ground-truth Generation . . . . . .
2.3 Datasets . . . . . . . . . . . . . . . . . . .
2.3.1 UPC Dataset . . . . . . . . . . . .
2.3.2 CESCA Dataset . . . . . . . . . . .
2.3.3 MAWI Dataset . . . . . . . . . . .
2.3.4 PAM Dataset . . . . . . . . . . . .
xix
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
2
4
7
9
9
9
10
12
13
14
14
15
16
17
18
18
20
21
22
xx
3 Traffic Classification with NetFlow
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . .
3.2 Methodology . . . . . . . . . . . . . . . . . . . . . .
3.2.1 Traffic Classification Method . . . . . . . . . .
3.2.2 Training Phase and Ground-truth . . . . . . .
3.2.3 Machine Learning and Validation Process . .
3.2.4 Performance Metrics . . . . . . . . . . . . . .
3.2.5 Evaluation Datasets . . . . . . . . . . . . . .
3.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3.1 Performance with Unsampled NetFlow . . . .
3.3.2 Performance with Sampled NetFlow . . . . . .
3.4 Analysis of the Sources of Inaccuracy under Sampling
3.4.1 Error in the Traffic Features . . . . . . . . . .
3.4.2 Changes in the Flow Size Distribution . . . .
3.4.3 Flow Splitting . . . . . . . . . . . . . . . . . .
3.5 Dealing with Sampled NetFlow . . . . . . . . . . . .
3.5.1 Improving the Classification Method . . . . .
3.5.2 Evaluation in Other Network Environments .
3.5.3 Lessons Learned and Limitations . . . . . . .
3.6 Related Work . . . . . . . . . . . . . . . . . . . . . .
3.7 Chapter Summary . . . . . . . . . . . . . . . . . . .
4 Autonomic Traffic Classification System
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . .
4.2 Traffic Classification System . . . . . . . . . . . . . .
4.2.1 The Application Identifier . . . . . . . . . . .
4.2.2 The Autonomic Retraining System . . . . . .
4.2.3 Training Dataset Generation . . . . . . . . . .
4.3 Evaluation . . . . . . . . . . . . . . . . . . . . . . . .
4.3.1 Evaluation of Labeling DPI-based Techniques
4.3.2 Training Dataset Evaluation . . . . . . . . . .
4.3.3 Retraining Evaluation . . . . . . . . . . . . .
4.3.4 Retraining Evaluation by Institution . . . . .
4.4 Related Work . . . . . . . . . . . . . . . . . . . . . .
4.5 Chapter Summary . . . . . . . . . . . . . . . . . . .
CONTENTS
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
25
25
26
27
27
29
30
30
31
32
34
36
37
41
44
47
47
49
50
52
53
.
.
.
.
.
.
.
.
.
.
.
.
55
55
56
57
59
61
62
62
64
66
70
73
73
CONTENTS
xxi
5 Streaming-based Traffic Classification
5.1 Introduction . . . . . . . . . . . . . . . . . . . . .
5.2 Classification of Evolving Network Data Streams .
5.2.1 Hoeffding Tree . . . . . . . . . . . . . . .
5.2.2 Hoeffding Adaptive Tree . . . . . . . . . .
5.2.3 Inputs of our System . . . . . . . . . . . .
5.3 Methodology . . . . . . . . . . . . . . . . . . . .
5.3.1 MOA: Massive Analysis Online . . . . . .
5.3.2 The MAWI Dataset . . . . . . . . . . . . .
5.4 Hoeffding Adaptive Tree Parametrization . . . . .
5.4.1 Numeric Estimator . . . . . . . . . . . . .
5.4.2 Grace Period . . . . . . . . . . . . . . . .
5.4.3 Tie Threshold . . . . . . . . . . . . . . . .
5.4.4 Split Criteria . . . . . . . . . . . . . . . .
5.4.5 Leaf Prediction . . . . . . . . . . . . . . .
5.4.6 Other Parameters . . . . . . . . . . . . . .
5.5 Hoeffding Adaptive Tree Evaluation . . . . . . . .
5.5.1 Single Training Evaluation . . . . . . . . .
5.5.2 Interleaved Chunk Evaluation . . . . . . .
5.5.3 Chunk Size Evaluation . . . . . . . . . . .
5.5.4 Periodic Training Evaluation . . . . . . . .
5.5.5 External Evaluation . . . . . . . . . . . .
5.6 Related Work . . . . . . . . . . . . . . . . . . . .
5.7 Chapter Summary . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
75
75
77
77
78
78
79
79
80
80
81
82
83
83
84
87
88
89
90
90
93
94
96
97
.
.
.
.
.
.
.
.
.
.
.
.
99
99
100
101
101
103
103
104
104
106
106
109
110
6 Validation of Traffic Classification Methods
6.1 Introduction . . . . . . . . . . . . . . . . . .
6.2 Methodology . . . . . . . . . . . . . . . . .
6.2.1 The Testbed . . . . . . . . . . . . . .
6.2.2 Selection of the Data . . . . . . . . .
6.2.3 Extracting the Data for Processing .
6.2.4 The Classification Process . . . . . .
6.2.5 The PAM Dataset . . . . . . . . . .
6.3 Performance Comparison . . . . . . . . . . .
6.3.1 Sub-classification of HTTP Traffic . .
6.4 Lessons Learned and Limitations . . . . . .
6.5 Related Work . . . . . . . . . . . . . . . . .
6.6 Chapter Summary . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
xxii
CONTENTS
7 Other Improvements to Traffic Classifiers
7.1 Efficient Profiled Flow Termination Timeouts
7.1.1 Introduction . . . . . . . . . . . . . . .
7.1.2 Dataset . . . . . . . . . . . . . . . . .
7.1.3 Methodology . . . . . . . . . . . . . .
7.1.4 Results . . . . . . . . . . . . . . . . . .
7.1.5 Related Work . . . . . . . . . . . . . .
7.1.6 Section Summary . . . . . . . . . . . .
7.2 Early Classification of Network Traffic . . . .
7.2.1 Introduction . . . . . . . . . . . . . . .
7.2.2 Methodology . . . . . . . . . . . . . .
7.2.3 Results . . . . . . . . . . . . . . . . . .
7.2.4 Related Work . . . . . . . . . . . . . .
7.2.5 Section Summary . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
111
. 112
. 112
. 113
. 113
. 114
. 123
. 124
. 126
. 126
. 127
. 130
. 139
. 140
8 Conclusions
141
8.1 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
Bibliography
158
Appendix A
159
8.2 UPC dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
8.3 PAM dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
Appendix B
8.4 Publications . . . . . . . . . . . . .
8.4.1 Journals . . . . . . . . . . .
8.4.2 Conferences . . . . . . . . .
8.4.3 Technical Reports . . . . . .
8.4.4 Supervised Master Students
8.4.5 Datasets . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
163
. 163
. 163
. 163
. 164
. 164
. 164
List of Figures
2.1
Traffic breakdown of the traces in the UPC dataset . . . . . . . . . 19
3.1
3.2
29
Overview of the Machine Learning and validation process . . . . . .
Precision (mean with 95% CI) by application group (per flow) of
our traffic classification method (C4.5) with different sampling rates
3.3 Recall (mean with 95% CI) by application group (per flow) of our
traffic classification method (C4.5) with different sampling rates . .
3.4 Overall accuracy (mean with 95% CI) of our traffic classification
method (C4.5) and a port-based technique as a function of the sampling rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.5 Overall accuracy when removing the error introduced by the inversion of the features (UPC-I trace, using UPC-II for training) . . . .
3.6 Validation of Eq. 3.11 against the empirical distribution of the original flow length detected with p = 0.1 (UPC-II trace). . . . . . . . .
3.7 Flow length distribution of the detected flows when using several
sampling probabilities. The figure was obtained combining Eq. 3.11
with Eq. 3.12 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.8 Amount of split flows as a function of the sampling probability p.
The figure shows both the empirical results (UPC-II trace) and the
analytical ones (Eq. 3.14) . . . . . . . . . . . . . . . . . . . . . . .
3.9 Overall accuracy (mean with 95% CI) of our traffic classification
method with a normal and sampled training set . . . . . . . . . . .
3.10 Precision (mean with 95% CI) by application group (per flow) of
our traffic classification method with a sampled training set . . . .
3.11 Recall (mean with 95% CI) by application group (per flow) of our
traffic classification method with a sampled training set . . . . . . .
3.12 Overall accuracy with a normal and sampled training set using the
traces from other environments . . . . . . . . . . . . . . . . . . . .
xxiii
32
33
35
42
43
44
46
48
48
49
51
xxiv
4.1
4.2
4.3
LIST OF FIGURES
4.5
Application Identifier and Autonomic Retraining System architecture
DPI labeling contribution in the CESCA dataset . . . . . . . . . . .
Impact of the Autonomic Retraining System on the Application
Identifier with the selected configuration (i.e., naive training policy
with 500K) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Comparative of the Autonomic Retraining System with other retraining solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Comparative of the Autonomic Retraining System by institution . .
70
71
5.1
5.2
5.3
5.4
5.5
5.6
5.7
5.8
5.9
5.10
5.11
5.12
Impact of the Numeric Estimator parameter on the HAT technique
Impact of the Grace Period parameter on the HAT technique . . .
Impact of the Tie Threshold parameter on the HAT technique . . .
Impact of the Split Criteria parameter on the HAT technique . . .
Impact of the Leaf Prediction parameter on the HAT technique . .
HAT comparison with single training configuration . . . . . . . . .
Interleaved Chunk evaluation with default configuration . . . . . . .
HAT accuracy comparison by chunk size . . . . . . . . . . . . . . .
HAT cost comparison by chunk size . . . . . . . . . . . . . . . . . .
HAT accumulated cost comparison by chunk size . . . . . . . . . .
Interleaved Chunk comparison with [61] configuration . . . . . . .
Interleaved Chunk evaluation with CESCA dataset . . . . . . . . .
82
84
85
86
87
90
91
92
93
94
95
96
7.1
Blocking scenario behavior. Both sides send data, but since there
is no acknowledgment from the other side they try to retransmit
(after a time which grows in every attempt) until finally break the
connection with a RST . . . . . . . . . . . . . . . . . . . . . . . . . 119
Termination proportions in ISP Core trace, green stands for the
standard FIN process, yellow for the unclosed flows, red for the
RST and intense red for the RST in a blocking scenario . . . . . . . 120
CDF of inter-packet times with RST termination in a blocking scenario121
Evaluation of PACE timeout in ISP Core trace by application . . . 124
Performance of PACE timeout in ISP Core trace by flows . . . . . . 125
Diagram of the Continuous Training Traffic Classification system
based on TIE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
K-dimensional tree evaluation without the support of the relevant
ports . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
K-dimensional tree evaluation with the support of the relevant ports136
K-dimensional tree evaluation with the support of the relevant ports137
4.4
7.2
7.3
7.4
7.5
7.6
7.7
7.8
7.9
56
63
69
LIST OF FIGURES
xxv
7.10 Accuracy by application group (seven packet sizes and selected list
of ports as parameters) . . . . . . . . . . . . . . . . . . . . . . . . . 138
xxvi
LIST OF FIGURES
List of Tables
2.1
2.2
2.3
2.4
2.5
2.6
IANA assigned port numbers for several well-known applications . .
Strings at the beginning of the payload of P2P protocols defined in [1]
Characteristics of the traces in the UPC dataset . . . . . . . . . . .
Traffic mix of the CESCA dataset . . . . . . . . . . . . . . . . . . .
Top 10 applications by flow in the MAWI dataset . . . . . . . . . .
Traffic mix of the PAM dataset . . . . . . . . . . . . . . . . . . . .
3.1
3.2
3.3
3.4
Set of 10 NetFlow-based features . . . . . . . . . . . . . . . . .
Application groups used by L7-filter . . . . . . . . . . . . . . . .
Elephant flow distribution in the UPC dataset . . . . . . . . . .
Overall accuracy of our traffic classification method (C4.5) and
port-based technique for the traces in our dataset . . . . . . . .
Average of the relative error of the flow features as a function of
(UPC-II trace) . . . . . . . . . . . . . . . . . . . . . . . . . . .
Characteristics of the traces from other environments . . . . . .
3.5
3.6
4.1
4.2
4.3
4.4
4.5
.
.
.
a
.
p
.
.
10
11
18
20
21
22
. 28
. 30
. 31
. 32
. 41
. 50
Features used by each classification technique in the Autonomic
Traffic Classification System . . . . . . . . . . . . . . . . . . . . . .
Application groups and traffic mix in the UPC-II and CESCA datasets
DPI techniques consumption in the Autonomic Retraining System .
Long-Term policy evaluation in the Autonomic Traffic Classification
System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Training dataset evaluation in the Autonomic Traffic Classification
System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
59
60
64
66
67
5.1
HAT parametrization for network traffic classification . . . . . . . . 89
6.1
6.2
DPI-based techniques evaluated . . . . . . . . . . . . . . . . . . . . 100
DPI evaluation by application . . . . . . . . . . . . . . . . . . . . . 107
xxvii
xxviii
6.3
6.4
6.5
7.1
7.2
7.3
7.4
7.5
7.6
7.7
7.8
7.9
7.10
7.11
7.12
LIST OF TABLES
HTTP sub-classification by nDPI . . . . . . . . . . . . . . . . . . . 108
FLASH evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
DPI summary results . . . . . . . . . . . . . . . . . . . . . . . . . . 109
Characteristics of the ISP traces in the evaluation dataset . . . .
Flow usage after the sanitization process . . . . . . . . . . . . . .
Traffic mix by flow in the evaluation dataset . . . . . . . . . . . .
PCK-PCK times TCP in ISP Core trace by application (ms) . .
PCK-PCK times TCP in ISP Mob trace by application (ms) . . .
PCK-PCK times UDP in ISP Core trace by application (ms) . .
DAT-DAT times TCP in ISP Core trace by application (ms) . . .
RST times in ISP Core trace by application (ms) . . . . . . . . .
Evaluation of TCP timeout by application (ms) . . . . . . . . . .
Evaluation of UDP timeout by application (ms) . . . . . . . . . .
Characteristics of the traffic traces in our dataset . . . . . . . . .
Speed Comparison (flows/s) : Nearest Neighbor vs K-Dimensional
Tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.13 Memory Comparison: Nearest Neighbor vs K-Dimensional Tree .
7.14 Building Time Comparative: Nearest Neighbor vs K-Dimensional
Tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.15 Evaluation of the Continuous Training system by training trace and
set of relevant ports . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
113
114
115
116
117
117
118
119
122
122
132
. 133
. 133
. 134
. 139
List of Acronyms
ADWIN Adaptive Sliding Window.
AFTP Anonymous File Transfer Protocol.
ANN Approximate Nearest Neighbor.
ARPANET Advanced Research Projects Agency Network.
BPF Berkeley Packet Filter.
CAIDA Center for Applied Internet Data Analysis.
CDF Cumulative Distribution Function.
CESCA Supercomputing Centre of Catalonia.
CI Confidence Interval.
COST European Cooperation in Science and Technology.
CPU Central Processing Unit.
DAT Data Packet.
DD Direct Download.
DNS Domain Name System.
DPI Deep Packet Instpection.
FN False Negative.
FP False Positive.
FTP File Transfer Protocol.
GGSN Gateway GPRS Support Node.
GMM Gaussian Mixture Model.
GPL General Public License.
GPRS General Packet Radio Service.
GTP GPRS Tunneling Protocol.
HAT Hoeffding Adaptive Tree.
xxix
xxx
HMM Hidden Markov Model.
HT Hoeffding Tree.
HTTP Hypertext Transfer Protocol.
IANA Internet Assigned Numbers Authority.
IM Instant Messaging.
IMAP Internet Message Access Protocol.
IOS Internetwork Operating System.
IP Internet Protocol.
IPFIX Internet Protocol Flow Information Export.
ISP Internet Service Provider.
IT Information Technology.
LX Linux.
MAC Media Access Control.
MAWI Measurement and Analysis on the WIDE Internet.
ML Machine Learning.
MOA Massive Online Analysis.
NBAR Network Based Application Recognition.
NETBIOS Network Basic Input/Output System.
NII National Institute of Informatics.
NN Nearest Neighbor.
NTP Network Time Protocol.
OSI Open Systems Interconnection.
P2P Peer to Peer.
PACE Protocol and Application Classification Engine.
PAM Passive and Active Measurement Conference.
PCAP Packet Capture.
PCK Packet.
POP3 Post Office Protocol v3.
QoS Quality of Service.
RDP Remote Desktop Protocol.
RTCP Real-Time Control Protocol.
RTMP Real-Time Messaging Protocol.
List of Acronyms
List of Acronyms
RTP Real-Time Transport Protocol.
RTSP Real-Time Streaming Protocol.
SGSN Serving GPRS Support Node.
SMTP Simple Mail Transfer Protocol.
SNMP Simple Network Management Protocol.
SSH Secure Shell.
SSL Secure Sockets Layer.
STD Standard Deviation.
SVM Support Vector Machine.
TCP Transmission Control Protocol.
TIE Traffic Identification Engine.
ToS Type of Service.
TP True Positive.
TSP Time Synchronization Protocol.
UDP User Datagram Protocol.
UPC Universitat Politècnica de Catalunya BarcelonaTech.
URL Uniform Resource Locator.
VBS Volunteer-Based System.
VFML Very Fast Machine Learning.
VM Virtual Machine.
VNC Virtual Network Computing.
W7 Windows 7.
WEB World Wide Web.
WEKA Waikato Environment for Knowledge Analysis.
WITS Waikato Internet Traffic Storage.
WoW World of Warcraft.
XP Windows XP.
xxxi
xxxii
List of Acronyms
Chapter 1
Introduction
The Internet is becoming central in our life and work. From chatting in Facebook
to discovering the cure for cancer, almost every aspect of our life is somehow related to the Internet. Long gone are the days when the original ARPANET was
born. Since then, the network has been in constant evolution transforming the
initial ARPANET in what we today know as the Internet, an enormous conglomerate of interconnected computer networks. Despite its important role in our life,
the knowledge about its operation is far from being completely understood. The
continuous introduction of new network architectures, protocols and applications
during the last decades resulted in an ever evolving entity difficult to study and
understand. This has spurred the research community to better analyze the network traffic and bring some light about the complex operation of the Internet.
Particularly, a new field of study, usually referred to as traffic classification, has
become crucial for understanding the Internet. The classification of network traffic not only satisfies our curiosity, but it also has many important applications for
network operators and IT administrators. BitTorrent, Skype (i.e., P2P), Youtube,
Netflix (i.e., streaming) or Megaupload (i.e., direct download) are some examples
of network applications that at some point completely changed the paradigms of
the Internet. The classification of network traffic helps in many different manners.
For instance, studying how new applications impact on the network can help to
better plan new infrastructures, architectures or protocols. An accurate classification can also help Internet Service Providers (ISPs) to apply reliable techniques
to apply Quality of Service (QoS) policies based on the needs of the applications
(e.g., VoIP calls). Finally, this opens a new range of billing possibilities for ISPs
to take profit of their infrastructures based on their actual usages.
The desire of network operators would be to be able to accurately classify all
1
2
CHAPTER 1. INTRODUCTION
the traffic of their networks online. However, the continuous evolution of Internet
applications and their techniques to avoid being detected make their identification
a very challenging task. Thus, the research community has thrown itself into the
search of techniques to accurately identify and classify the traffic. Nevertheless,
a wide range of unaddressed functional problems arise when those techniques are
applied in real scenarios dealing with tremendous amount of traffic and limited
resources. In this thesis, we address the network traffic classification problem from
a more practical point of view, emphasizing those problems that arise when this
classification is performed online in operational networks.
The rest of this chapter discusses the motivations and challenges behind this
dissertation and also provides an overview of the thesis and its main contributions.
Finally, it describes the structure of this manuscript.
1.1
Motivations
The network traffic classification problem can be seen as a never ending race between application developers, on the one side, and the network operators and the
research community, on the other. Originally, the most common and simplest
technique to identify network applications was based on the port numbers (e.g.,
those registered by the IANA [2]). This solution was very efficient and relatively
accurate with traditional applications. The arrival of bandwidth eater P2P-based
applications (e.g., Napster, eMule) was the first turning point. Network operators usually reacted to the high consume of those applications directly blocking
their default used ports [3]. To avoid being blocked, application developers and
users replied by using dynamic ports or even using registered ones from traditional
applications, making the solution based on the well-known ports completely unreliable. Nowadays it is widely accepted that this method is no longer valid due to
the inaccuracy and incompleteness of its classification results [1, 4–6].
The response from the research community was the introduction of Deep Packet
Inspection (DPI) techniques for the classification of those applications [1, 5–11].
DPI methods are based on searching for characteristic signatures (or patterns) in
the packet payloads. This solution is potentially very accurate and it is commonly
used as ground-truth generator for the validation of other techniques. However,
its application online is computationally extremely expensive. Once again, the
applications reacted by implementing protocol encryption (i.e., cipher the payload)
to evade classification.
Machine Learning (ML) techniques were later proposed as a promising solution
1.1. MOTIVATIONS
3
to the well-known limitations of port- and DPI-based techniques [12–36]. Nguyen
et al. survey and compare the literature in the field of ML-based traffic classification in [37]. In general terms, most ML methods study in an offline phase the
relation between a pre-defined set of traffic features (e.g., port numbers, flow sizes,
inter-arrival times) and each application. This set of features is used to build a
model, which is later used to identify the network traffic online. In order to partially decrease the precision of ML techniques, network applications implemented
protocol obfuscation methods that randomly modify the characteristics of their
traffic.
In parallel with ML and DPI techniques, other less prolific approaches were also
proposed. An interesting alternative following a different approach are the hostbehavior-based methods [38–40]. BLINC [38] is arguably the most well-known
exponent of this alternative branch. These methods base the classification on the
behavior of the end-hosts that produced the traffic. Another solution similar to
the well-known ports are the IP-based techniques [41, 42]. This approach uses the
knowledge extracted from previously seen IP addresses from well-known applications (e.g., Youtube) to classify the traffic related to those applications.
Although most of the existing solutions can achieve high accuracy in theory,
there is no universal method suitable for every possible network scenario. In addition, their deployment in production networks, with a large number of users
and connections, presents practical constraints that existing methods do not completely address. As a result, currently existing methods have reached limited
success among network operators and managers. From our point of view, three
main aspects should be addressed in order to introduce new techniques that are
feasible and reliable for network traffic classification in production networks, which
are the basis of this thesis.
The first issue is related to the deployment of the classification solutions. The
limited resources and high throughput requirements of production networks usually do not allow the extra burden imposed by existing techniques. For example,
DPI techniques usually require access to the payload of each packet to search for
the set of patterns. This operation needs expensive hardware in order to cope with
the high data rates of nowadays networks [10, 43, 44]. In contrast, ML methods
need to extract for each flow (i.e., connection) the set of features used as input
for the classification. Although less demanding, some features are very costly to
extract and require specific hardware to compute them [45]. In addition, network
operators usually apply packet sampling in their monitoring solutions in order not
to compromise the correct operation of the network in highly consuming situations (e.g., attacks). However, the impact of packet sampling on the proposed
4
CHAPTER 1. INTRODUCTION
classification techniques remains unknown.
The second issue is the high maintenance costs involved in current classification
solutions. For instance, most ML techniques [36, 37, 46] rely on a costly training
phase that requires human intervention. As shown by Li et al. in [47], these
techniques usually need regular updates in order to adapt to new traffic or new
networks. This not only implies the involvement of the network operator, but also
a specific knowledge for carrying out the task. Similarly to ML techniques, DPI
and IP-based techniques require regular updates of the signature and IP base used
for the classification.
And finally, the third issue is related to the impossibility of validating and
comparing the different proposals. Most traffic classification solutions proposed in
the literature report very high accuracy. However, most solutions base their results
on a private ground-truth (i.e., dataset), usually labeled by techniques of unknown
reliability (e.g., ports-based or DPI-based techniques [46,48–50]). That makes the
comparison and validation of different proposals very difficult in order to decide
what solution is more suitable for each scenario. The use of private datasets is
derived from the lack of publicly available datasets with payload. Mainly because
of privacy issues, researchers are not allowed to share their datasets with the
research community. Another crucial problem is the reliability of the techniques
used to set the ground-truth. Most papers show that researchers usually obtain
their ground-truth through port-based or DPI-based techniques [46, 48–50]. The
poor reliability of port-based techniques is already well known [1, 51]. According
to conventional wisdom DPI-based techniques are, in principle, one of the most
accurate techniques. However, their actual reliability is still unknown.
These three aspects hinder the success of currently existing techniques in production networks. Therefore, network operators usually rely on obsolete solutions
to classify their network traffic [52, 53] or are just able to classify a small portion
of their network traffic due to the expensive hardware, and complex deployment
and maintenance costs of state-of-the-art classification solutions.
1.2
Contributions and Impact
This thesis is about filling the gap between the real-world requirements from the
network industry, and the research being carried out in the network traffic classification field. The contributions of this thesis follow the same fronts of the
motivations. Most of the techniques for network traffic classification proposed in
the state of the art reported high accuracy, however those works usually lack to
1.2. CONTRIBUTIONS AND IMPACT
5
address the practical issues related to the introduction of those systems in real
environments.
Ease of deployment is a crucial element of a realistic network traffic classification solution. We address this problem in Chapter 3. As described, most
of the state-of-the-art solutions need additional (usually expensive) hardware in
order to perform the classification [10, 43–45]. To address this issue this thesis focuses on the use of NetFlow as input for the classification. NetFlow is a
widely extended protocol developed by Cisco to export IP flow information from
routers and switches [54]. The main challenge when using NetFlow is the limited
amount of information available, which complicates the classification [15, 46], but
significantly reduces the cost of the solution and allows its rapid deployment in
production networks given that most network devices already support NetFlow or
one of its variants (e.g., IPFIX [55]). However, due to the limited resources in core
networks, NetFlow is usually operated under packet sampling (i.e., only processing
a sample of the traffic). Chapter 3 studies the impact of Sampled NetFlow on the
classification. Using the C4.5 decision tree ML technique, we study empirically
and theoretically how packet sampling affects the classification with NetFlow. In
particular, we identify three different sources of error when using packet sampling
in NetFlow: (i) error in the estimation of the flow features, (ii) changes in the flow
size distribution, and (iii) splitting of sparse flows. Afterwards, we propose a simple but effective methodology to improve the classification accuracy of ML-based
techniques in this scenario. As a result, our ML-based classification solution is
able to accurately classify the traffic only using as input the information provided
by Sampled NetFlow. This considerably facilitates the deployment of our proposal
given that the input can be easily provided by the current network infrastructure
without compromising its correct operation.
The maintenance is the other practical problem barely addressed by the stateof-the-art proposals. As already pointed out, the continuous evolution of the Internet traffic forces classification solutions to be regularly updated to adapt to
traffic changes. Usually, proposed solutions avoid this limitation and only focus
on obtaining a high classification accuracy with a static dataset. This forces network operators to intuitively decide how often it is necessary to manually perform
the tedious update of each instance of the classification solution. Unlike the rest
of the literature, Chapter 4 focuses on this issue with the goal of automatically
maintaining the classification solution without human intervention while keeping
high classification accuracy. With the objective of proposing a realistic solution
we present a technique that combines different classification methods only using
NetFlow as input. Furthermore, we design an autonomic retraining system that is
6
CHAPTER 1. INTRODUCTION
able to keep the classifier accurate along time. All these features (i.e., high accuracy, ease of deployment and maintenance) make the proposed system a realistic
solution for network traffic classification in backbone networks.
Chapter 5 introduces the use of stream-based ML techniques for network traffic
classification. These techniques, scarcely used in this field [56–58], have very appealing features for network traffic classification: (i) process an instance at a time
and inspects it only once, (ii) use a predefined amount of memory, (iii) work in a
bounded amount of time and (iv) are ready to predict at any time. We propose
a technique based on Hoeffding Adaptive Trees, a stream-based decision tree that
can be as accurate as batch techniques like the already mentioned C4.5. To extract sound conclusions, we evaluated this technique with the MAWI dataset [59],
a 13 years long dataset from a transatlantic link in Japan. Our technique is able
to automatically adapt to the changes of the traffic with just a small sample of
labeled data, making our solution very easy to maintain. Also, following previous
recommendations, we are able to accurately classify the traffic using only NetFlow
data making our solution very easy to deploy.
The literature has mainly focused on proposing techniques to achieve high
accuracy while we, in addition, focused on making them feasible in real scenarios. However, an essential element that remained unaddressed also hindered the
adoption of new techniques, which is the validation and comparison of proposed
techniques. The wide range of techniques, datasets and ground-truth generators
make the comparison and validation of new proposed techniques a very difficult
task. Chapter 6 sheds light to this problem by evaluating the reliability of six wellknown DPI-based techniques usually used for ground-truth generation. We created
a labeled dataset and used Volunteer-Based System (VBS) [60] to guarantee the
accurate labeling process. The obtained results help researchers to understand the
reliability and implications of using the different DPI-based techniques (i.e., PACE,
OpenDPI, nDPI, Libprotoident, L7-filter and NBAR) for ground-truth generation.
In addition, as part of this research, we published some of the datasets used
in our evaluations in an attempt to allow for the fair comparison and validation
of state-of-the-art techniques. The dataset used in Chapter 3 anonymized and
labeled with L7-filter has been shared with multiple institutions as can be seen in
the Appendix A. The dataset created in Chapter 6 had special impact, because it is
the first reliable labeled dataset with full packet payloads available to the research
community. To the best of our knowledge, this is the only dataset available that
makes it possible the comparison and validation of ML and DPI based techniques.
Although these are the most important contributions of this thesis, other improvements related to the network classification field have been also achieved.
1.3. THESIS ORGANIZATION
7
Almost every proposed solution for network traffic classification needs to keep
track of the connections (i.e., flows) to properly classify the traffic. Usually it is
necessary to store information of every packet to extract the features or payload
later used in the classification. This process is of capital importance in a scenario
with limited resources like core networks. As a result, a proper mechanism for
expiration of the flows is a key point in the optimization of this costly process.
Section 7.1 presents the study of an efficient profiled termination timeout to optimize this operation. First, we carefully study the flow termination by group of
applications with current Internet traffic. From that characterization, we went
further and proposed an expiration technique to achieve optimized timeouts in a
profiled way.
Finally, the last contribution of this thesis proposes an early classification technique able to classify the traffic just using the size of the first packets of the flows.
We revisited in Section 7.2 the viability of using the Nearest Neighbor algorithm
(NN) for online traffic classification, which has been often discarded in previous
studies due to its poor classification speed. In order to address this well-known
limitation, we presented an efficient implementation of the NN algorithm based
on a K-dimensional tree data structure. Our results show that our method can
achieve very high accuracy by looking only at the first packet of a flow. When
the number of analyzed packets is increased to seven, the accuracy of our method
increases beyond 95%. This early classification feature is very important, since it
allows network operators to quickly react to the classification results.
1.3
Thesis Organization
This thesis dissertation is organized as follows. Chapter 2 presents important
background related to the network traffic classification field and describes in detail
the datasets used throughout this manuscript. Chapter 3 first studies the impact of using NetFlow and packet sampling for network traffic classification both
theoretically and empirically. Afterwards, it presents a solution to improve the
accuracy of ML-based techniques using Sampled NetFlow. This work that aims
to facilitate the deployment of state-of-the-art techniques is based on the journal
paper published in [46]. Next, Chapter 4 proposes a complete and feasible solution
for network traffic classification in backbone networks. Using the previous knowledge and combining different literature techniques, the proposed system is able to
sustain a high classification accuracy along time without human intervention and
just using Sampled NetFlow for the classification. The solution possesses the main
8
CHAPTER 1. INTRODUCTION
requirements for an appealing network classification system (i.e., easy to deploy
and maintain, high classification accuracy and completeness). Its design, operation and evaluation have been published in [61]. Chapter 5 introduces the use of
stream-based ML techniques for network traffic classification, proposing a solution
based on Hoeffding Adaptive Trees. The nature of the network scenarios makes
this technique the perfect candidate for a continuous traffic classification solution.
Chapter 6, based on the work published in [62], presents a first step to address the
validation problem of the network traffic classification field. To achieve this goal,
we created and published a reliable labeled dataset and compared the accuracy of
several well-known DPI-based techniques used in the literature for ground-truth
generation. Chapter 7 presents two additional contributions in the network traffic
classification field. First, we propose a methodology based on profiled termination
timeouts to efficiently expire flows in network monitoring solutions. Afterwards, a
technique based on K-dimensional trees is proposed for early network traffic classification. These works are based on the publications in [63, 64]. Finally, Chapter 8
concludes this thesis and presents some ideas for future work.
Chapter 2
Background
This chapter presents some important aspects related to network traffic classification, which are necessary to understand the contributions of this thesis. Also, it
describes in detail the datasets used throughout this manuscript.
2.1
Network Traffic Classification Approaches
This section describes the state-of-the-art approaches in the field of network traffic
classification. We refer to this classification as the identification of network applications (e.g., Skype, eMule) or group of applications (e.g., P2P, WEB) at Layer 7
of the OSI model. The application identification problem has been changing due
the efforts of two opponents that are in a continuous competition. On the one
hand, the applications, and especially those that do not want to be detected (e.g.,
P2P applications), in order to use the network resources without control. On the
other hand, network operators, researchers and even ISPs who need to know the
traffic characteristics of their networks to manage the resources or even charge
the users based on their consumption. This rivalry yielded the creation of several
techniques for network traffic classification.
2.1.1
Port-based Approach
As mentioned in Section 1.1, traditionally, the classification of network applications
has been carried out using the well-known ports technique. At the beginning, the
initial network applications registered their ports in the IANA [2]. The solution of
the well-known ports identifies the traffic according to the ports registered in the
9
10
CHAPTER 2. BACKGROUND
IANA. Table 2.1 shows an example of several ports assigned by the IANA with the
corresponding application. For example, web applications use port 80 and email
applications use port 25 (SMTP) to send emails and port 110 (POP3) to receive
them.
Assigned Port
20
21
22
23
25
53
80
110
123
161
3724
Application
FTP Data
FTP Control
SSH
Telnet
SMTP
DNS
HTTP
POP3
NTP
SNMP
WoW
Table 2.1: IANA assigned port numbers for several well-known applications
This method is no longer valid because of the inaccuracy and incompleteness of
its classification results [4–6]. It is difficult to set bounds on the current accuracy
of this method because it mainly depends on the characteristics of the network
being monitored and the system to establish the ground-truth used to validate
them. There are some studies where this technique achieves an accuracy of 50%70% [6, 16], while in others, like in the present work, the accuracy can be less
than 20%. However, the complete literature agrees that the well-known ports
technique is not able to classify the new generation of applications that are the
ones that consume more bandwidth. This new type of applications, especially P2P
applications, use different strategies to camouflage their traffic in order to evade
detection. Additionally, they use dynamic ports in their connections or ports from
other well-known applications (e.g., 80 typically used by web applications).
2.1.2
Payload-based Approach
The first alternative to the well-known ports method was the inspection of the
packet payloads to identify the network traffic [1,4–8,10,11,50,65]. These methods,
usually called DPI techniques, examine the content of the packets looking for
2.1. NETWORK TRAFFIC CLASSIFICATION APPROACHES
11
characteristic signatures of the applications in the traffic. The work done under
this approach aims to identify specially P2P applications because they are the most
assiduous to use camouflage strategies and evade the detection of the well-known
ports technique. Table 2.2 presents an example of patterns used by Karagiannis
et al. in [1].
P2P Protocol
eDonkey 2000
Fasttrack
BitTorrent
Gnutella
Ares
String
0xe319010000
0xe53f010000
”Get /.hash”
0x2700000002980
”0x13Bit”
”GNUT” ”GIV”
”GND”
”GET hash:”
”Get sha1:”
Trans. Prot.
TCP/UDP
TCP
UDP
TCP
TCP
UDP
TCP
Table 2.2: Strings at the beginning of the payload of P2P protocols defined in [1]
Also, there are some papers [1, 4, 5] that presented hybrid methods using the
well-known ports method to identify the traditional network applications and a
deep packet inspection to classify the new ones. Although solutions based on pattern matching could achieve high accuracy, they had some problems with three
main factors. First, pattern searching in the payload of every packet implies a
high consume of resources that needs extremely expensive dedicated hardware to
handle nowadays networks. Second, it does not work with encrypted traffic, an increasing trend among the P2P applications. And third, a continuous update of the
set of application signatures is necessary to classify new versions and applications.
Although there are some proposals that tried to facilitate this process [14], the generation of new signatures is still a tedious and challenging task that requires human
supervision. Given the high resource requirements for the pattern searching, the
limitations with encrypted traffic and their continuous update requirement made
their application impractical in current high-speed networks. However, this family
of techniques is still the used solution to establish the ground-truth [8, 10, 11, 50].
12
2.1.3
CHAPTER 2. BACKGROUND
Flow features-based Approach
In order to overcome the limitations of the payload methods and find an alternative to identify applications without the necessity of packet inspection, Machine
Learning techniques were introduced. These methods detect, in an offline phase,
characteristic patterns of the different applications based on a set of features. More
in detail, Machine Learning methods use a (usually labeled) dataset from which a
set of features is extracted. This information serves as input of the ML technique
that extracts and outputs the knowledge in different structures (e.g., decision tree,
rules, clusters) depending on the ML technique used. The structure obtained is
later used to classify unlabeled instances assuming that the features of the still
unknown instances will have the same behavior as the known one.
The wide related work in this field can be structured in two main areas: the
supervised and unsupervised learning approaches. Next subsections provide brief
descriptions of these two areas. However, Nguyen et al. surveyed the literature
of the ML techniques in the field of application identification in [37] comparing in
detail all the different approaches.
Supervised Learning Approach
Supervised methods, also known as classification methods, extract knowledge
structures to classify new instances in pre-defined classes. It is important to
note that is called supervised because the output classes are pre-defined. The
process of a supervised ML methods start with a training dataset TS defined as,
T S = < x1 , y1 >, < x2 , y1 >, ..., < xN , yM >, where xi is the vector of values of
the features corresponding to the ith instance, and yi is its output class value. It
discovers the different relations between the instances and outputs a structure,
usually a decision tree or classification rules, that will classify the instances in a
discrete set y1 , y2 , ..., yM .
There is a lot of related work that use supervised techniques [12–19] with a
promising results.
Unsupervised Learning Approach
Unlike the supervised ML methods, unsupervised methods, also known as clustering methods, do not need a complete labeled training dataset. Therefore, the
output of the ML training does not classify instances in predefined classes. It is the
own method that discovers the natural clusters (groups) in the data. Similar to
the supervised methods, the clustering methods have been deeply studied [20–30]
2.1. NETWORK TRAFFIC CLASSIFICATION APPROACHES
13
Basically, three main clustering methods have been used in the network traffic classification literature: the classic k-means algorithm that forms clusters in
numeric domains, partitioning instances into disjoint clusters; the incremental
clustering that generates a hierarchical grouping of instances and; the probabilitybased method that assigns instances to classes probabilistically, not deterministically. In general, they achieved a slightly lower accuracy than supervised techniques, but with a much more lightweight training phase.
It is important to note that usually the clustering methods have a tradeoff
between the number of clusters of the output and the final accuracy. Having more
clusters you could classify better but the training and classification time of the
method will also increase.
Given their appealing characteristics, the ML-based traffic classification techniques are the most prolific solution in the literature. However, similarly to pattern matching techniques, ML methods should be periodically updated to adapt
the classification model to new applications. This process is difficult and usually
requires human supervision. Furthermore, the extraction of the set of features
used for the classification can require dedicated hardware to handle nowadays networks [45].
2.1.4
Host-behavior-based Approach
Another branch that appears to solve the limitations of the payload-based methods
is the host-behavior-based techniques. Karagiannis et al. with BLINC [38,40] and
Xu et al. in [39] developed a classification approach based on the behavior of the
hosts. The Karagiannis’ approach studies the behavior of the hosts at three levels:
• At social level, they value the popularity of a host depending on the number
of communications it has with different hosts.
• At functional level, they value the role of the host in the network finding
out if the host is a provider or consumer of a service, or it participates
in collaborative communications. For example, a host which has a lot of
connections at the same port is likely to be provider of a service in that port.
• At application level, they capture the transport layer interactions between
hosts trying to identify the application of origin. For example, a host with a
lot of connections from the same source port and IP to a unique destination
IP but different ports would be the behavior of a port scan attack.
14
CHAPTER 2. BACKGROUND
According to these three metrics and some refining heuristics, BLINC classifies
the behavior of the applications. BLINC classifies approximately 80%-90% of
the total number of flows with 95% accuracy. However, BLINC presents some
limitations regarding the necessity to work with traces collected in the edge and
with bidirectional flows. Furthermore, BLINC classifies the applications based on
behavior groups but it is not able to identify exactly the applications. It could
suppose a problem with applications that are theoretically from different groups
but with similar behavior (e.g., SopCast and BitTorrent).
2.1.5
IP-based Approach
IP addresses can carry important information related to the traffic they produce.
Following the same approach as the port-based techniques, these methods use the
information provided by IP addresses to classify the traffic. Two main solutions
to use the IP information have been proposed.
The first technique is based on the solely use of well-known IP addresses to
classify the traffic [41]. Basically, this technique tracks down in an offline phase the
IP addresses belonging to famous web sites (e.g., Google, Twitter). For instance,
the IP 31.13.83.16 belongs to Facebook, so the traffic coming or going to that IP
is directly classified as Facebook.
The second technique is also known as Service-based method [42]. A service is
defined as the triplet <IP, Port, Protocol> assigned to a specific application. The
list of services is also created in an offline phase using a dataset of labeled flows.
For example, Yoon et al. in [42] aggregate all the available flows by their triplet
and then, a service is created when a triplet has a minimum number of flows during
a specific time threshold and there is a predominant label for all these flows.
These techniques are very accurate, however its completeness has been significantly degraded given the migration of some applications to Content Delivery
Networks. Furthermore, and similarly to other approaches, periodically updates
should be applied in order to maintain the techniques accurate.
2.1.6
Comparing Approaches
So far, we presented a spread work in the field of techniques to identify applications.
However, it is not clear which one is the best method to identify applications
in network traffic. Some researchers addressed this problem in several papers
comparing different techniques [31–36].
2.2. NETWORK TRAFFIC CLASSIFICATION EVALUATION
15
In 2006, Williams et al. made in [33] perform evaluation between seven supervised algorithms. The used algorithms were C4.5 Decision Tree, Naive Bayes,
Nearest Neighbor, Naive Bayes Tree, Multilayer Perception Network, Sequential
Minimal Optimization and Bayesian Networks. They concluded that C4.5 is the
most accurate supervised technique with and accuracy of 99,4%, followed by Bayes
Network 99,32% and Naive Bayes Tree and Naive Bayes with 98,3%. Although
the accuracy difference is minimum, Williams et al. show significantly differences
regarding the classification time, where C4.5 outperformed the rest of methods
being twice faster than Naive Bayes and more than five times faster than Bayes
Network.
Also in 2006, Erman et al. presented in [31] the study of the comparison of the
supervised method Naive Bayes and the unsupervised method AutoClass. They
concluded that the unsupervised method achieved a 91% accuracy, 9% more than
the supervised method.
At the end of 2008, Kim et al. presented in [36] a comparison of different
supervised methods, BLINC and the well-known port method implemented by
CoralReef [52]. Comparing the different approach the supervised ML techniques
achieved better accuracy than BLINC and CoralReef. Among the different supervised methods (i.e., Naive Bayes, Naive Bayes Kernel Estimation, Bayesian
Network, C4.5, k-NN, Neural Networks and Support Vector Machine (SVM)) the
paper concluded that SVM outperformed the rest of supervised methods with more
than 98% overall accuracy while C4.5, in the same scenario, achieved an overall
accuracy of ≈94%. However, C4.5 classified the instances a hundred times faster
than SVM.
We can conclude that according to the literature there is not a clear technique
that outperforms the rest of techniques. It seems that all the methods could achieve
a very good accuracy in specific scenarios. However, as mentioned in Section 1.1
and further discussed in Chapter 6, the comparison of techniques of the literature
is limited by the use of private datasets usually labeled by techniques of unknown
reliability.
2.2
Network Traffic Classification Evaluation
This section describes some important aspects related to the evaluation of the
results in the network traffic classification field. Two main elements are common
in all the evaluations of network traffic classification solutions. The first element
are the metrics used in order to measure the quality of the proposals. The second
16
CHAPTER 2. BACKGROUND
element is the ground-truth used as reference in order to validate and compute the
metrics.
2.2.1
Performance Metrics
Depending on the requirements, this thesis uses several metrics in order to asses
the quality of the experiments. The most extended metric used in the network
traffic literature is the accuracy. The accuracy is computed by calculating the
number of correctly classified flows. The exact definition of the accuracy metric
in a non binary classification would be:
PN
i=1 (T P )
PN
i=1 (F P )
i=1 (T P ) +
Accuracy = PN
where:
• N: number of categories (e.g., groups of applications).
• TP (True Positives): The number of correctly identified flows for a specific
category.
• FP (False Positives): The number of falsely identified flows for a specific
category.
Other metrics usually computed are the precision (i.e., positive predictive
value) and the recall (i.e., sensitivity) by category. The precision is defined as
TP
. This metric indicates the fraction of instances from a particular category
T P +F P
that are correctly classified from the total amount of instances classified as that
category.
P
On the other hand, the recall of a particular category is defined as T PT+F
,
N
where FN (False Negatives) stands for those flows of the category under study
that are classified as belonging to another category. This metric indicates the
fraction of instances from a particular category that are correctly classified from
the total amount of instances of that category.
Although these metrics are the most popular metrics used in the network traffic
classification literature, they have some limitations. Because of this, sometimes it
is necessary to compute other metrics to confirm the results obtained. The Kappa
coefficient metric is considered to be more robust because it takes into account the
2.2. NETWORK TRAFFIC CLASSIFICATION EVALUATION
17
correct classifications occurring by chance. We computed the Kappa coefficient as
explained by Cohen in [66]:
Po − Pe
k=
1 − Pe
being:
PN
i=1 (T P )
PN
i=1 (T P ) +
i=1 (F P )
P o = PN
N
X
Pe =
(Pi1 × Pi2 )
i=1
where:
• Pi1 : proportion of apparition of the category i for the observer 1.
• Pi2 : proportion of apparition of the category i for the observer 2.
The Kappa coefficient takes values close to 0 if the classification is mainly due to
chance agreement. On the other hand, if the classification is due the discriminative
power of the classification technique then the values are close to 1.
Another important metric is the completeness. This metric measures the
amount of traffic not classified of the dataset. That is, the percentage of traffic that remains unknown after the classification.
2.2.2
Ground-truth Generation
The establishment of the ground-truth is one of the most critical phases of any
network traffic classification evaluation since the entire classification process relies
on the reliability of the technique used to obtain the first labeling used as reference.
Three main approaches are used in the literature.
The most accurate technique could be the human manual labeling of the
dataset. However, the manual labeling flow by flow is a very tedious task making
it impracticable with the current datasets containing millions of flows.
Most papers show that researchers usually obtain their ground-truth through
port-based or DPI-based techniques [46, 48–50]. However, the poor reliability of
port-based techniques is already well known, given the use of dynamic ports or
well-known ports of other applications [1,51]. Because of this, the use of port-based
techniques is no longer recommended for the generation of the ground-truth.
18
CHAPTER 2. BACKGROUND
Table 2.3: Characteristics of the traces in the UPC dataset
Name
UPC-I
UPC-II
UPC-III
UPC-IV
UPC-V
UPC-VI
UPC-VII
Date
11-12-08
11-12-08
12-12-08
12-12-08
21-12-08
22-12-08
10-03-09
Day
Thu
Thu
Fri
Fri
Sun
Mon
Tue
Start Time
10:00
12:00
16:00
18:30
16:00
12:30
03:00
Duration
15 min
15 min
15 min
15 min
1h
1h
1h
Flows
2 985 K
3 369 K
3 369 K
3 020 K
7 146 K
9 718 K
5 510 K
Bytes
53 G
63 G
55 G
48 G
123 G
256 G
78 G
Avg. Util
482 Mbps
573 Mbps
500 Mbps
436 Mbps
279 Mbps
582 Mbps
177 Mbps
The DPI-based techniques are currently the most used solutions for groundtruth generation. Some examples, widely used in the literature, are L7-filter [8],
OpenDPI [9], PACE [10], nDPI [11] and Libprotoident [50]. These techniques
usually combine pattern matching and heuristics to, according to conventional
wisdom, accurately label the traffic. Nevertheless, the reliability of DPI-based
techniques is still unknown.
2.3
Datasets
This chapter describes the datasets used in the different works of this thesis. Most
of them have been published in order to allow further comparison and validation
of our proposals.
2.3.1
UPC Dataset
The UPC dataset consists of seven full-payload traces collected at the Gigabit
access link of the Universitat Politècnica de Catalunya BarcelonaTech (UPC),
which connects about 25 faculties and 40 departments (geographically distributed
in 10 campuses) to the Internet through the Spanish Research and Education
network (RedIRIS). This link provides Internet access to about 50 000 users. The
traces were collected at different days and hours trying to cover as much diverse
traffic from different applications as possible. Table 2.3 presents the details of the
traces of this dataset.
To set the ground-truth of the UPC dataset we used L7-filter [8]. This DPIbased technique presents overmatching problems with some of its patterns. In
2.3. DATASETS
19
order to reduce the inaccuracy of L7-filter we used three rules for the labeling:
• We apply the patterns in a priority order depending on the degree of overmatching of each pattern (e.g., skype patterns are in the latest positions of
the rule list).
• We do not label those packets that do not agree with the rules given by
pattern creators (e.g., packets detected as NTP with a size different than 48
bytes are not labeled).
• In the case of multiple matches, we label the flow with the application with
more priority, based on the quality of each pattern reported in the L7-filter
documentation. If the quality of the patterns is equal, the label with more
occurrences is chosen.
As already proposed in previous works (e.g., [20]), we also perform a sanitization process in order to remove incorrect or incomplete flows that may confuse
or bias the dataset. The sanitization process removes those TCP flows that are
not properly formed (e.g., without TCP establishment or termination, and flows
with packet loss or with out-of-order packets) from the training set. However, no
sanitization process is applied to UDP traffic.
Figure 2.1: Traffic breakdown of the traces in the UPC dataset
Figure 2.1 plots the application breakdown according to L7-filter of the traces in
the UPC dataset. The breakdown reflects the academic nature of the monitored
environment, with a large portion of HTTP and network traffic. We can also
observe that L7-filter is not able to classify about 20% of the traffic (labeled as
20
CHAPTER 2. BACKGROUND
unknown). This was a common problem of the tools available to establish the
ground-truth, as already reported in previous works (e.g., [67]).
The set of seven labeled traces presented in Table 2.3 is publicly available in
an anonymized form to the research community at [68].
2.3.2
CESCA Dataset
The CESCA trace is a fourteen-days packet trace collected on February 2011 in
the 10-Gigabit access link of the Anella Cientı́fica, which connects the Catalan
Research and Education Network with the Spanish Research and Education Network. This link provides connection to Internet to more than 90 institutions. This
trace was collected with 1/400 flow sampling. We applied this ratio because it was
the lowest that allowed us to collect the trace without packet loss in our hardware.
To obtain a reliable ground-truth, we use a set of DPI-based techniques, including
PACE, a commercial DPI library provided by ipoque [10]. PACE is known to
have high accuracy with low false positive ratio. Moreover, to increase the completeness of the DPI classification we added two extra libraries, OpenDPI [9] and
L7-filter [8]. Table 2.4 describes the final traffic distribution of CESCA dataset.
Table 2.4: Traffic mix of the CESCA dataset
Group
Web
DD
Multimedia
P2P
Mail
Bulk
VoIP
DNS
Chat
Games
Encryption
Others
Applications
HTTP
E.g., Megaupload, MediaFire
E.g., Flash, Spotify, Sopcast
E.g., BitTorrent, eDonkey
E.g., IMAP, POP3
E.g., FTP, AFTP
E.g., Skype, Viber
DNS
E.g., Jabber, MSN Messenger
E.g., Steam, WoW
E.g., SSL, OpenVPN
E.g., Citrix, VNC
Flows
17 198 K
40 K
1 126 K
4 851 K
753 K
27 K
3 385 K
15 863 K
196 K
14 K
3 440 K
2 437 K
2.3. DATASETS
2.3.3
21
MAWI Dataset
The publicly available MAWI dataset [59] has unique characteristics to study
stream oriented techniques for network traffic classification. The MAWI dataset
consists of 15-minutes traces daily collected in a transit link since 2001 (i.e., 13
years). Although it is a static dataset, its long duration and amount of data makes
it the perfect candidate for stream-based evaluations. Furthermore, its duration
allows the study of the ability of classification techniques to adapt to the evolution
of the traffic.
Table 2.5: Top 10 applications by flow in the MAWI dataset
Year
Top 1
Top 2
Top 3
Top 4
Top 5
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
HTTP (49.44%)
HTTP (41.30%)
HTTP (30.22%)
HTTP (38.77%)
HTTP (31.02%)
DNS (33.34%)
DNS (50.42%)
DNS (50.82%)
DNS (44.31%)
DNS (48.67%)
DNS (39.91%)
DNS (44.93%)
DNS (54.87%)
DNS (42.11%)
DNS (37.75%)
DNS (22.55%)
DNS (26.45%)
DNS (30.80%)
HTTP (31.51%)
HTTP (31.61%)
HTTP (26.52%)
HTTP (22.04%)
HTTP (26.75%)
HTTP (29.55%)
HTTP (31.30%)
HTTP (26.78%)
DEMONWARE (3.27%)
OPASERV (11.81%)
OPASERV (22.46%)
SQL EXPLOIT (12.11%)
SQL EXPLOIT (13.85%)
SQL EXPLOIT (11.43%)
BITTORRENT (3.82%)
BITTORRENT (5.27%)
BITTORRENT (20.50%)
BITTORRENT (9.82%)
BITTORRENT (13.48%)
BITTORRENT (11.11%)
BITTORRENT (6.33%)
SMTP (2.37%)
DEMONWARE (4.16%)
SQL EXPLOIT (19.47%)
OPASERV (10.46%)
SKYPE (8.09%)
SKYPE (6.28%)
SKYPE (3.37%)
SKYPE (4.13%)
SKYPE (4.27%)
TEREDO (4.29%)
SKYPE (5.48%)
TEREDO (4.17%)
NTP (5.16%)
FTP (0.52%)
SMTP (1.79%)
SMTP (1.87%)
SMTP (3.40%)
MSN (3.91%)
BITTORRENT (4.39%)
SMTP (2.81%)
SQL EXPLOIT (3.86%)
GNUTELLA (2.74%)
SKYPE (3.76%)
TEREDO (4.30%)
SKYPE (2.12%)
SIP (1.27%)
Year
Top 6
Top 7
Top 8
Top 9
Top 10
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
NETBIOS (0.43%)
EMULE (0.62%)
EMULE (1.22%)
MSN (2.74%)
OPASERV (3.11%)
SMTP (2.66%)
SQL EXPLOIT (2.74%)
SMTP (3.39%)
SQL EXPLOIT (1.40%)
SSH (1.89%)
NTP (2.30%)
SSH (1.50%)
SKYPE (1.18%)
GNUTELLA (0.37%)
FTP (0.48%)
FTP (0.27%)
SKYPE (1.76%)
SMTP (2.41%)
OPASERV (1.73%)
SSH (1.67%)
SSH (2.04%)
SMTP (1.7%)
SMTP (1.17%)
SSH (1.01%)
NTP (1.31%)
SSH (1.11%)
CALL OF DUTY (0.28%)
GNUTELLA (0.43%)
NORTON (0.23%)
NETBIOS (1.07%)
BITTORRENT (2.10%)
MSN (1.66%)
MSN (0.84%)
MSN (1.61%)
SSH (0.83%)
SQL EXPLOIT (0.68%)
SMTP (0.59)
SIP (0.56%)
PANDO (0.93%)
HALF LIFE (0.22%)
MSN (0.23%)
GNUTELLA (0.2%)
GNUTELLA (0.51%)
TDS (1.21%)
PPLIVE (1.60%)
FTP (0.37%)
QQ (0.26%)
EMULE (0.76%)
SIP (0.48%)
EMULE (0.58%)
SMTP (0.44%)
SMTP (0.47%)
IRC (0.19%)
IRC (0.21%)
MSN (0.18%)
FTP (0.30%)
SMB (0.42%)
SMB (0.58%)
EMULE (0.34%)
ORBIT (0.24%)
PPSTREAM (0.32%)
NTP (0.41%)
SIP (0.43%)
CANON BJNP (0.36%)
CANON BJNP (0.33%)
To set the ground-truth of the MAWI dataset, we used a DPI technique. The
packets in the private version of this dataset are truncated after 96-bytes, which
considerably limits the amount of information available for the DPI techniques.
Because of this constraint, we rely our ground-truth labeling on Libprotoident [50].
The most important feature of Libprotoident is that its patterns are found just
in the first 4 bytes of payload of each direction of the traffic. Unexpectedly, that
data is enough to achieve very high accuracy classification as shown in [50, 62].
However, the MAWI dataset is characterized to have asymmetric traffic that can
reduce the effectiveness of Libprotoident. We performed a sanitization process and
focused on the TCP and UDP traffic from the MAWI dataset. Table 2.5 presents
22
CHAPTER 2. BACKGROUND
Table 2.6: Traffic mix of the PAM dataset
Application
eDonkey
BitTorrent
FTP
DNS
NTP
RDP
NETBIOS
SSH
Browser HTTP
Browser RTMP
Unclassified
Flows
176 K
62 K
876
6K
27 K
132 K
9K
26 K
46 K
427
771 K
Bytes
2 823.88 M
2 621.37 M
3 089.06 M
1.74 M
4.03 M
13 218.47 M
5.17 M
91.80 M
5 757.32 M
3 026.57 M
5 907.15 M
the top ten applications by flow along the thirteen years once the sanitization is
applied.
After the labeling and the sanitization process, the MAWI dataset consists of
almost 4 billions of unidirectional labeled flows. To the best of our knowledge, this
is the first work in the network traffic classification field that deals with this large
amount of data, which is necessary to extract sound conclusions from stream-based
evaluations.
2.3.4
PAM Dataset
The PAM dataset contains 1 262 022 flows captured during 66 days, between February 25, 2013 and May 1, 2013, which account for 35.69 GB of pure packet data. To
collect and accurately label the flows, we adapted Volunteer-Based System (VBS)
developed at Aalborg University [60]. The task of VBS is to collect information
about Internet traffic flows (i.e., start time of the flow, number of packets contained
by the flow, local and remote IP addresses, local and remote ports, transport layer
protocol) together with detailed information about each packet (i.e., direction,
size, TCP flags, and relative timestamp to the previous packet in the flow). For
each flow, the system also collects the process name associated with that flow.
The process name is obtained from the system sockets. This way, we can ensure
the application associated to a particular traffic. Additionally, the system collects
some information about the HTTP content type (e.g., text/html, video/x-flv ). The
2.3. DATASETS
23
captured information is transmitted to the VBS server, which stores the data in a
MySQL database. The design of VBS was initially described in [60].
Thanks to this methodology, the application name tag was present for 520 993
flows (41.28 % of all the flows), which account for 32.33 GB (90.59 %) of the data
volume. Additionally, 14 445 flows (1.14 % of all the flows), accounting for 0.28 GB
(0.78 %) of data volume, could be identified based on the HTTP content-type field
extracted from the packets. Therefore, we were able to successfully establish the
ground-truth for 535 438 flows (42.43 % of all the flows), accounting for 32.61 GB
(91.37 %) of data volume. The remaining flows are unlabeled due to their short
lifetime (below ∼1 s), which made VBS incapable to reliably establish the corresponding sockets. However, we have included the unlabeled flows in the dataset in
order to keep the consistency of the traffic. The classes together with the number
of flows and the data volume are shown in Table 2.6.
We have published this labeled dataset with full packet payloads in [68]. Therefore, it can be used by the research community as a reference benchmark for the
validation and comparison of network traffic classifiers.
24
CHAPTER 2. BACKGROUND
Chapter 3
Traffic Classification with
NetFlow
3.1
Introduction
This chapter addresses the problem of the deployment of traffic classification solutions. As described in Chapter 1, the limited resources and high throughput
requirements of production networks together with the extended application of
packet sampling hinder the deployment of state-of-the-art classification techniques.
Particularly, concerning the existing ML-based traffic classification solutions, it is
worth noting that:
1. Most ML-based techniques only operate with packet-level traces, which requires the deployment of additional (often expensive) monitoring hardware [45],
increasing the burden on network operators.
2. The impact of sampling in traffic classification still remains unknown, although traffic sampling (e.g., Sampled NetFlow) is a common practice among
network operators.
3. Most ML-based techniques rely on a costly and time-consuming training
phase, which often involves manual inspection of a large number of connections.
In order to counteract the three aforementioned issues, in this chapter:
25
26
CHAPTER 3. TRAFFIC CLASSIFICATION WITH NETFLOW
• We study the classification problem using NetFlow data instead of packetlevel traces (open issue 1). NetFlow is a widely extended protocol developed
by Cisco to export IP flow information from routers and switches [54]. The
main challenge when using NetFlow is the limited amount of information
available to be used as features of ML classification methods.
• We use the well-known C4.5 decision tree technique in order to analyze the
impact of traffic sampling on the classification accuracy with Sampled NetFlow (open issue 2). The main limitation in this case is the low sampling rates
typically used by network operators (e.g., 1/1 000) to allow routers to handle
worst-case traffic scenarios and network attacks. We analyze this problem,
both empirically and using sampling theory, and find that packet sampling
has a severe impact on the performance of the classification method.
• We propose an automatic ML solution that does not rely on any human
intervention and significantly reduces the impact of traffic sampling on the
classification accuracy (open issues 2 and 3). The main novelty of our solution is that our training set consists of sampled instances that can better
capture the properties of the sampled traffic that will be processed in the
classification phase. In contrast, previous works used the same training set
for both sampled and unsampled scenarios.
• We evaluate the performance of the method using a wide set of traffic
traces collected in a large research and education network (4 hours) and
validate the results using traces from different environments (CAIDA [69]
and WITS [70]).
The remainder of this chapter is organized as follows. Section 3.2 describes
the NetFlow-based methodology. Section 3.3 presents an experimental evaluation
of the traffic classification method with Sampled NetFlow. Section 3.4 provides a
theoretical analysis of the impact of sampling using sampling theory. Section 3.5
proposes a variant of the classification technique to reduce the impact of sampling.
Section 3.6 reviews in greater detail the related work. Finally, Section 3.7 concludes
the chapter.
3.2
Methodology
This section describes a new ML-based classification method for Sampled NetFlow
and its automatic training phase. We also present the methodology and datasets
3.2. METHODOLOGY
27
used for its evaluation and testing.
3.2.1
Traffic Classification Method
We decided to use the C4.5 supervised ML method [71] given its high accuracy
and low overhead compared to other ML techniques as will be further discussed in
Section 3.6. The C4.5 algorithm builds in an offline phase a decision tree from a
set of pre-classified examples (i.e., training set). The training set consists of pairs
<flow, application>, where the flow is represented as a vector of features and the
application is a label that identifies the network application that generated the
flow. In our case, the training set contains real traffic flows collected at a large
university network (described in Section 3.2.5), and the vector of features includes
those features of a flow that are relevant to predict its application.
The main difference between our method and those proposed in previous works
is that, when using Sampled NetFlow data, the potential number of features is significantly reduced due to the limited amount of information available in NetFlow
version 5 records (i.e., the most extended version). Based on this basic information, we also compute some additional features, such as the average packet size or
average inter-arrival time, resulting in 10 features in total. Table 3.1 presents the
set of features used in this work. Unlike in some previous works (e.g., [15]), we do
not use information about the IP addresses. This decision can negatively impact
the final accuracy of the classification method but it allows us to build a classification model less dependent on the particular network scenario used in the training
phase. In addition, in the presence of sampling, we correct some features (e.g.,
number of packets and bytes) by multiplying their sampled values by the inverse
of the sampling rate. Table 3.1 shows how each feature is computed according
to the sampling rate (p) being applied. In Section 3.4, we provide a theoretical
analysis of the error in the estimated features under sampling.
3.2.2
Training Phase and Ground-truth
One of the main limitations of existing ML methods is that they require a very
expensive training phase, which usually involves manual inspection of a potentially
very large number of connections. In order to automatize the training phase, we developed an automatic training tool based on L7-filter. L7-filter [8] is an open source
DPI tool that looks for characteristic patterns in the packet payloads and labels
them with the corresponding application. This computationally expensive process
is possible during the training phase given that it is executed offline. Finally, the
28
CHAPTER 3. TRAFFIC CLASSIFICATION WITH NETFLOW
Table 3.1: Set of 10 NetFlow-based features
Feature
sport
dport
protocol
T oS
f lags
duration
packets
bytes
pkt size
iat
Description
Source port of the flow
Destination port of the flow
IP protocol value
Type of Service from the first packet
Cumulative OR of TCP flags
Duration of the flow in nsec precision
Total number of packets in the flow
Flow length in bytes
Average packet size of the flow
Average packet inter-arrival time
Value
16 bits
16 bits
8 bits
8 bits
6 bits
tsend − tsini
packets
p
bytes
p
bytes
packets
duration
packets/p
training phase ends with the generation of a C4.5 decision tree as described in
Section 3.2.1. Although the training phase must be performed using packet-level
traces, only the NetFlow traffic features listed in Table 3.1 are considered to build
the decision tree and to evaluate its performance.
The establishment of the ground-truth is one of the most critical phases of any
ML traffic classification method, since the entire classification process relies on the
accuracy of the first labeling. Although L7-filter is probably not as accurate as the
manual inspection methods used in previous works, it is automatic and does not
require human intervention. This is a very important feature given that manual
inspection is often not possible in an operational network. For example, each of
the traces used in this section contains more than 2 million flows (see Table 2.3).
However, our method can also be trained using manually labeled datasets if more
accurate results are needed or to avoid the limitations of DPI techniques.
The complete operation to obtain the ground-truth is described in Section 2.3.1.
Because of the limitations of L7-filter, we removed the flows labeled as unknown
from the training set, assuming that they have similar characteristics to other
known flows. This assumption is similar to that of ML clustering methods, where
unlabeled instances are classified according to their proximity in the feature space
to those that are known. More sophisticated labeling methods have recently appeared [9–11, 50, 67], which could be also used to reduce the number of unknown
flows.
3.2. METHODOLOGY
29
Training Phase
Validation Phase
Online Classification
Packet
Traces
Labelling
process
NetFlow v5
parser
NetFlow
enabled
router
NetFlow
feature
extraction
Model building
(WEKA)
Classification
model
(C4.5)
Classification
output
Figure 3.1: Overview of the Machine Learning and validation process
3.2.3
Machine Learning and Validation Process
Figure 3.1 summarizes the training phase presented in Section 3.2.2 and the validation process used in the evaluation. First, we use a labeled dataset to train
our ML method. This dataset establishes the ground-truth of our system and is
obtained with L7-filter as described in Section 3.2.2. Next, the NetFlow feature extraction module extracts, from each flow in the dataset, the set of NetFlow-based
features listed in Table 3.1. After applying the sanitization process described in
Section 3.2.2, we obtain the training set. This training set is then processed by the
Model building module to generate the C4.5 decision tree that will be used in the
validation phase (classification model ). We use the WEKA ML software suite [72]
to build the J48 decision tree, an open source java implementation of the C4.5
(release 8) technique.
In the validation phase, we evaluate the performance of the classification model
obtained in the training phase. We included the resulting decision tree in the
SMARTxAC network monitoring system [73] and processed a large set of traces
described later in Section 3.2.5. Although these traces also contain the application
label, this information is only relevant to validate the accuracy of our method.
After extracting the NetFlow-based features from the evaluation traces similarly to
NetFlow, we classify each flow using the C4.5 decision tree generated in the training
phase. The quality of the model is measured by comparing the output of the
decision tree and the label obtained with L7-filter, according to the performance
metrics presented in Section 3.2.4.
30
CHAPTER 3. TRAFFIC CLASSIFICATION WITH NETFLOW
Table 3.2: Application groups used by L7-filter
Group
P2P
HTTP
VoIP
Network
Streaming
DNS
Others
Chat
Email
FTP
Games
3.2.4
Applications
File-sharing applications based on peer-to-peer architecture (e.g., Bittorrent, eDonkey)
HyperText Transfer Protocol applications (e.g, http, httpcachemiss, httpcachehit)
Voice communications applications (e.g., Skype, TeamSpeak, Ventrilo)
Network applications (e.g, BGP, DHCP, Netbios, SSH, Telnet, VNC,
X11)
Stream media applications (e.g., Shoutcast, PPLive, Itunes, QuickTime)
Domain Name System traffic
CVS, Hddtemp, IPP, LPD, Subversion, TSP...
Instant messaging applications (e.g, Aim, IRC, Jabber, MSN Messenger,
Yahoo Messenger)
Email traffic (e.g, IMAP, POP3, SMTP)
File transfer applications (e.g., FTP, Gopher, Tftp, UucP)
Massively multiplayer online games (e.g, Battlefield, Counter-Strike
Source, Day of Defeat Source, WoW)
Performance Metrics
In the evaluation, we use three representative performance metrics: overall accuracy, precision, and recall. Section 2.2.1 describes in detail how we computed these
metrics. In order to simplify the exposition of the results, we show all performance
results broken down by application group, according to the groups presented in
Table 3.2, which are based on those defined in the L7-filter documentation [8].
A flow is defined as the 5-tuple consisting of the source and destination IP
addresses, ports, and protocol. In the evaluation, we also present the accuracy of
the classification method per packets and bytes.
3.2.5
Evaluation Datasets
In this evaluation, we used the UPC dataset described in Section 2.3.1. This
dataset consists of seven full-payload traces collected at the Gigabit access link of
the Universitat Politècnica de Catalunya BarcelonaTech (UPC). The traces were
collected at different days and hours trying to cover as much diverse traffic from
different applications as possible.
Table 3.3 shows the estimations of the percentage of elephant flows and the
traffic they generate. As we will show later in Section 3.3, this parameter has a
3.3. RESULTS
31
Table 3.3: Elephant flow distribution in the UPC dataset
Name
UPC-I
UPC-II
UPC-III
UPC-IV
UPC-V
UPC-VI
UPC-VII
Flows
2 985
3 369
3 474
3 020
7 146
9 718
5 510
K
K
K
K
K
K
K
Elephant Flows
% Flows % Bytes
0.035818% 52.17%
0.048619% 61.45%
0.041587% 59.58%
0.048149% 59.79%
0.014151% 66.08%
0.042271% 54.51%
0.014075% 72.44%
significant impact on our classification results. In order to identify elephant flows,
we use the metric proposed in [74], where a flow is considered as elephant when
its size is greater than the mean flow size observed in the trace plus three times
the standard deviation.
Among the different traces in the UPC dataset, we selected a single trace (UPCII) for the training phase, which is the one that contains the highest diversity in
terms of instances from different applications. We limit our training set to one
trace in order to leave a meaningful number of traces for the evaluation that are
not used to build the classification model.
As mentioned in Section 2.3.1, the set of seven labeled traces is publicly available in an anonymized form to the research community at [68].
3.3
Results
We first evaluate the performance of the traffic classification method with unsampled NetFlow. These results are then used as a reference to analyze the impact
of traffic sampling on our method. In all experiments, we use the UPC-II trace
in the training phase, as described in Section 3.2.5, and evaluate the accuracy of
our method with the seven traces presented in Table 2.3, using the performance
metrics described in Section 3.2.4.
32
CHAPTER 3. TRAFFIC CLASSIFICATION WITH NETFLOW
Table 3.4: Overall accuracy of our traffic classification method (C4.5) and a portbased technique for the traces in our dataset
Name
UPC-I
UPC-II
UPC-III
UPC-IV
UPC-V
UPC-VI
UPC-VII
Flows
89.17%
93.67%
90.77%
91.12%
89.72%
88.89%
90.75%
Overall
C4.5
Packets
66.37%
82.04%
67.78%
72.58%
70.21%
68.48%
61.37%
Accuracy
Port-based
Bytes
Flows
56.53%
11.05%
77.97%
11.68%
61.80%
9.18%
63.69%
9.84%
61.21%
6.49%
60.08%
16.98%
40.93%
3.55%
Figure 3.2: Precision (mean with 95% CI) by application group (per flow) of our
traffic classification method (C4.5) with different sampling rates
3.3.1
Performance with Unsampled NetFlow
Table 3.4 presents the overall accuracy (in flows, packets and bytes) broken down
by trace without applying traffic sampling. Compared to the related work [12–36],
our traffic classification method obtains similar accuracy to previous packet-based
ML techniques, despite the fact that we are only using the features provided by
NetFlow.
The C4.5 decision tree obtained in the training phase classifies in average 22 057
flows/second with unsampled data using a 3GHz machine with 4GB of RAM. This
high classification speed is also reported in other works [32–34, 36].
While the overall flow accuracy is 90.59% in average, the classification accuracy
in packets and bytes is, as in the related work, significantly lower (69.83% and
3.3. RESULTS
33
Figure 3.3: Recall (mean with 95% CI) by application group (per flow) of our
traffic classification method (C4.5) with different sampling rates
60.32%, respectively). This is a common phenomenon with supervised learning
due to the heavy-tailed nature of current Internet traffic. This results in much
more instances of mice flows in the training set than of elephant flows, which
hinders the identification of a small number of flows that carry a large fraction of
the traffic. Table 2.3 confirms that, although the percentage of elephant flows is
less than 0.1%, they represent more than 50% of the traffic in our traces. This
explains the lower performance achieved in terms of packets and bytes compared to
that obtained in terms of flows. However, this effect can be solved by artificially
increasing the number of instances of elephant flows in the training set, as we
discuss in Section 3.5.3.
Table 3.4 shows that the overall accuracy is similar for all traces. As expected,
results for the UPC-II trace are slightly better than for the rest, given that this
trace was the one used in the training phase. However, the accuracy per byte with
the UPC-VII trace is significantly lower than with the other traces, since this trace
was collected at night, when the aggregated traffic mix was more different than
that of the training set (e.g., backup and file transfer applications). An example
of this different traffic profile can be observed in Table 2.3, which shows that the
percentage of elephant flows for those traces collected during weekends (UPC-V)
or at night (UPC-VII) is significantly lower (about 0.01%) but, in contrast, they
carry a much larger fraction of the traffic (between 60-70%).
Table 3.4 also compares the overall accuracy of our method to a simple technique based on the destination port numbers [2]. As expected, even when using
only NetFlow-compatible features, the accuracy of supervised learning methods
is significantly higher than that of port-based techniques. It is also important to
note that, while in previous works the accuracy of port-based techniques was in
34
CHAPTER 3. TRAFFIC CLASSIFICATION WITH NETFLOW
the 50-70% range [15], the accuracy of this method in our traces is only around
10%, which shows the important differences between the traces used in different
studies.
Figure 3.2 presents the precision by application group. When sampling is not
applied (p = 1), the precision is around 90% for most application groups. Figure 3.3 also shows similar results for the recall by application group. Nevertheless,
the performance for some applications (e.g., Streaming, Chat and Others) is significantly lower because L7-filter detected very few flows of these applications in the
training set (see Figure 2.1, UPC-II trace). As a result, this inaccuracy has a very
limited impact on the overall performance of our method. In order to improve the
classification accuracy for these groups, we could always include more instances of
these applications in the training set, as we further discuss in Section 3.5.3.
3.3.2
Performance with Sampled NetFlow
So far, we have shown that our traffic classification method can achieve similar accuracy than previous packet-based ML techniques [31,36], but using only NetFlow
data. Next, we study the impact of packet sampling on our classification method
using Sampled NetFlow.
Figure 3.4a shows that the flow classification accuracy degrades drastically with
sampling, decreasing from 90.59%, when all packets are analyzed, until 51.02%,
when only one packet out of 1 000 is selected. The accuracy decreases much faster
for high sampling rates and stabilizes (and even slightly increases) after p = 10%.
This effect is inherent to packet sampling, since mice flows tend to rapidly disappear with packet sampling, while the percentage of elephant flows remains fairly
stable at low sampling rates. For example, we checked that the percentage of
elephant flows in our traces increases by two orders of magnitude (from around
0.05% to 5%) at low sampling rates. Recall that NetFlow only supports static
packet sampling and network operators tend to set the sampling rate to very low
values in order to allow routers to sustain worst-case traffic mixes and network
attacks.
The accuracy of the port-based technique remains almost constant in the presence of sampling (about 10-30%). The accuracy per flow somewhat increases with
the sampling rate, which indicates that large flows are easier to identify using the
port numbers than small flows. In contrast, the accuracy per packet and byte tends
to slightly decay compared to that per flow because, while the percentage of flows
stabilizes at low sampling rates, the percentage of packets and bytes decreases in
the same proportion as the sampling rate.
3.3. RESULTS
35
(a) Accuracy by flows
(b) Accuracy by packets
(c) Accuracy by bytes
Figure 3.4: Overall accuracy (mean with 95% CI) of our traffic classification
method (C4.5) and a port-based technique as a function of the sampling rate
Figures 3.4b and 3.4c also show that our method is more resilient to sampling
for large flows than for small ones, since the accuracy per packet and byte degrades
more slowly with the sampling rate than per flow. On the one hand, this is because
missing a single packet of a small flow always has a larger impact in the estimation
of the flow features than in larger flows. On the other hand, as previously discussed,
while our method classifies small flows better than larger ones, small flows tend to
rapidly disappear in the presence of sampling.
Figure 3.2 presents the flow precision by application group at different sampling
rates. The precision for most applications, except for those with very few instances
in the training set (e.g., Streaming, Chat and Others), decreases with the same
36
CHAPTER 3. TRAFFIC CLASSIFICATION WITH NETFLOW
proportion depicted in Figure 3.4. Figure 3.3 also shows similar results for the
recall by application group.
Both figures when interpreted together give us important information about
the quality of our classification method. For example, it can be observed that the
precision and recall for HTTP is around 70% and 35% (with p = 0.01) respectively,
which indicates that while 70% of the flows classified as HTTP are actually HTTP
traffic, more than 65% of the HTTP flows in the dataset are incorrectly classified
as belonging to another application. On the contrary, while the precision for FTP
applications is relatively low in the presence of sampling (< 50%), the recall is very
high (> 90%), which means that more than 90% of the FTP flows in the dataset
are correctly classified. However, more than 50% of flows identified as FTP belong
to other applications.
According to the results presented in this section, we can conclude that the
impact of traffic sampling on supervised learning solutions for traffic classification
is severe. We consider that an average accuracy of about 50% for sampling rates
below 10% is unacceptable to most network operators. Surprisingly, to the best of
our knowledge, this is the first work that shows such a drastic impact of sampling
in the field of traffic classification. The only previous work that studied the traffic
classification problem in the presence of sampling [15] reported instead a minimal
decrease of accuracy. We attribute the different results to the more complete
datasets used in this study (4 hours of a gigabit link), which are harder to classify
using only the port numbers. As we later discuss in Section 3.4, the port numbers
are resilient to sampling, which explains the differences in the results. As in this
section, a complete, unsampled dataset was employed to train the system in [15].
In contrast, Section 3.5 shows that by simply using a sampled training set to
build the classification model, the accuracy and recall figures can be significantly
improved.
3.4
Analysis of the Sources of Inaccuracy under
Sampling
In Section 3.3.2, we have shown the large impact of packet sampling on the accuracy of a supervised learning technique for traffic classification. However, packet
sampling can affect the classification accuracy in many different ways. In particular, we have detected three different sources of error when using packet sampling
in NetFlow: (i) error in the estimation of the flow features, (ii) changes in the flow
3.4. ANALYSIS OF THE SOURCES OF INACCURACY UNDER SAMPLING37
size distribution, and (iii) splitting of sparse flows. In this section, we analyze
these sources of inaccuracy both theoretically and empirically.
3.4.1
Error in the Traffic Features
One of the main sources of inaccuracy under sampling is the estimation of the
traffic features. Most of the per-flow features provided by Sampled NetFlow cannot
be directly used by the classifier. Instead, they must be inverted prior to the
classification process in order to be comparable to the values observed in the
training set. We present a theoretical analysis, using sampling theory, of the error
in the estimation of the features presented in Table 3.1. We show that, although
most of the features are fairly accurate for large flows and moderate sampling
rates, some features are clearly biased. For those features that are unbiased, we
find that the variance of the error is significant for small flows and low sampling
rates. We also present experimental evidence that indicates that this source of
error contributes most to the results presented in Section 3.3.2.
• Source port, destination port, and protocol. These are the only features in Table 3.1 that are not affected by packet sampling, because they
have exactly the same value in all the packets of a flow. Thus, their inversion has no impact in the classification accuracy. If one or more packets of
the flow are sampled, then the error is 0, while if the flow is not sampled at
all it does not contribute to the error. The probability of missing a flow with
n packets is (1 − p)n , where p is the sampling rate.
• Flags and ToS. Under random sampling, the probability p of sampling
a packet is independent from the other packets. Let m be the number of
packets of a particular flow with the flag f set (i.e., f = 1), where f ∈ {0, 1}.
The probability of incorrectly estimating the value of f under sampling is
(1 − p)m , independently of how the packets with the flag set are distributed
over the flow. The expected value of the absolute error is:
E[f − fˆ] = f − E[fˆ] = f − (1 − (1 − p)m ) = f − 1 + (1 − p)m
(3.1)
Eq. 3.1 shows that fˆ is biased, since the expectation of the error is (1 − p)m
when f = 1, and it is only 0 when f = 0. That is, with packet sampling,
fˆ tends to underestimate f , especially when f = 1 and m or p are small.
For example, if we have a flow with 100 packets with the flag ACK set
38
CHAPTER 3. TRAFFIC CLASSIFICATION WITH NETFLOW
(m = 100) and p = 1%, the expectation of the error in the flag ACK is
(1 − 0.01)100 ≈ 0.37. The flag SYN and the ToS are particular cases, where
we are only interested in the first packet and, therefore, m ∈ {0, 1}.
• Number of packets. With sampling probability p, the number of sampled
packets x from a flow of n packets follows a binomial distribution x ∼ B(n, p).
Thus, the expected value of the estimated feature n̂ = x/p is:
E[n̂] = E
1
1
= E[x] = np = n
p
p
p
hxi
(3.2)
which shows that n̂ is an unbiased estimator of n (i.e., the expected value of
the error is 0). The variance of n̂ is:
hxi
1
1
1
Var[n̂] = Var
= 2 Var[x] = 2 np(1 − p) = n(1 − p)
p
p
p
p
(3.3)
Hence, the variance of the relative error can be expressed as:
h
h n̂ i
n̂ i
1
1
1−p
Var 1 −
= Var
= 2 Var[n̂] = 2 n(1 − p) =
n
n
n
np
np
(3.4)
Eq. 3.4 indicates that, for a given p, the variance of the error decreases with
n. That is, the variance of the error for elephant flows is smaller than for
smaller flows. The variance also increases when p is small. For example, with
1−0.01
p = 1%, the variance of the error of a flow with 100 packets is 100×0.01
= 0.99,
which is not negligible.
P
• Flow size. The original size b of a flow is defined as b = ni=1 bi , where n is
the total number of packets of the flow and bi is the size of each individual
packet. Under random sampling, we can estimate b from a subset of sampled
packets by renormalizing their size:
b̂ =
n
X
i=1
wi
bi
p
(3.5)
where wi ∈ {0, 1} are Bernoulli distributed random variables with probability
p. We can show that b̂ is an unbiased estimator of b, since E[b̂] = b:
3.4. ANALYSIS OF THE SOURCES OF INACCURACY UNDER SAMPLING39
n
X
n
n
1 X
1X
bi
= E
w i bi =
E[wi bi ] =
E[b̂] = E
wi
p
p i=1
p i=1
i=1
n
n
n
1X
1X
1 X
=
bi E[wi ] =
bi p = p
bi = b
p i=1
p i=1
p i=1
(3.6)
The variance of b̂ is obtained as follows:
X
X
n
n
n
1
bi
1 X
= 2 Var
Var[wi bi ] =
Var[b̂] = Var
wi
w i bi = 2
p
p
p i=1
i=1
i=1
=
n
n
n
1 X 2
1 X 2
1−pX 2
b
Var[w
]
=
b
p(1
−
p)
=
bi
i
p2 i=1 i
p2 i=1 i
p i=1
(3.7)
Thus, the variance of the relative error is:
n
h
h b̂ i
b̂ i
1
1 1−pX 2
Var 1 −
= Var
= 2 Var[b̂] = 2
bi =
b
b
b
b p i=1
Pn
bi2
1−p
(3.8)
=
Pi=1
p ( ni=1 bi)2
P
P
which decreases with n, since ni bi2 ≤ ( ni bi)2 . This indicates that the
variance of the error can be significant for small sampling rates and short
flows. For example, if we have a flow with 100 packets of 1500 bytes each,
100×15002
the variance of the error with p = 1% is 1−0.01
× (100×1500)
2 = 0.99.
0.01
• Duration and interarrival time. The flow duration is defined as d =
tn − t1 , where t1 and tn are the timestamps of the first and last packets of
the original flow. Under sampling, this duration is estimated as dˆ = tb − ta ,
where ta and tb are the timestamps of the first and last sampled packets
respectively. Thus, the expected value of dˆ is:
ˆ = E[tb − ta ]
E[d]
= E[tb ] − E[ta ]
n
a
X
X
= E tn −
iati − E t1 +
iati
i=1
i=b
X
X
n
a
= (tn − t1 ) − E
iati + E
iati
i=b
i=1
(3.9)
40
CHAPTER 3. TRAFFIC CLASSIFICATION WITH NETFLOW
where iati is the interarrival time between packets i and i − 1, and a is a
random variable that denotes the number of missed packets until the first
packet of the flow is sampled (i.e., the number of packets between t1 and ta ).
Therefore, the variable a follows a geometric distribution with probability
p, whose expectation is 1/p. By symmetry, we can consider the number of
packets between b and n to follow the same geometric distribution. In this
case, we can rewrite Eq. 3.9 as follows:
ˆ = (tn − t1 ) − ((E[n − b] E[iat]) + (E[a] E[iat]))
E[d]
iat = (tn − t1 ) − 2
p
(3.10)
where iat is the average interarrival time of the non-sampled packets. Eq. 3.10
ˆ > 0). In other
shows that the estimated duration is biased (i.e., E[d − d]
ˆ
words, d always underestimates d. The bias is (2 × iat/p), if we consider
the average interarrival time to be equal between packets 1 . . . a and b . . . n.
c to correct this bias, because this feaHowever, we cannot use the feature iat
ˆ In fact, Eq. 3.10 indicates that the feature
ture is obtained directly from d.
ˆ
c is also biased, since iat
c = d/n̂.
iat
In order to illustrate with empirical data the impact of packet sampling on
the estimated features, we present in Table 3.5 the average of the relative error
per flow of the features in the UPC-II trace as a function of the sampling rate.
We can observe that the error for almost all the features is significant, and it is
particularly large for n̂ and b̂. These large errors are explained by the fact that the
average number of packets per flow in the UPC-II trace is only 6.67, given that
in NetFlow flows are unidirectional and the trace contains a large portion of DNS
traffic (i.e., single-packet flows). According to Eq. 3.4, the variance of the error
for small flows is very large. For example, with p = 0.01 and n = 1, the variance
= 99. The same observation holds
of the error in the number of packets is 1−0.01
0.01
for the flow size.
Given this large error, one would expect the error in the flow features to be the
main source of inaccuracy in the classification results presented in Section 3.3.2.
In order to confirm this intuition, we performed an experiment to quantify the
error introduced by the inverted features in the final classification accuracy. We
replaced the estimated features of each flow in the UPC-I trace by their original
values, which were directly obtained from the unsampled trace. Note that this
cannot be done in the general case, because under sampling the original features
3.4. ANALYSIS OF THE SOURCES OF INACCURACY UNDER SAMPLING41
Table 3.5: Average of the relative error of the flow features as a function of p
(UPC-II trace)
Feature
sport
dport
proto
fˆ
dˆ
n̂
b̂
c
iat
p = 0.5
0.00
0.00
0.00
0.05
0.22
0.66
0.76
0.29
p = 0.1
0.00
0.00
0.00
0.16
0.60
3.66
3.86
0.65
p = 0.05
0.00
0.00
0.00
0.18
0.66
6.90
7.05
0.71
p = 0.01
0.00
0.00
0.00
0.22
0.77
29.69
29.71
0.78
p = 0.005
0.00
0.00
0.00
0.23
0.79
55.17
55.09
0.80
p = 0.001
0.00
0.00
0.00
0.24
0.81
234.61
234.24
0.82
are unknown. Figure 3.5 shows the accuracy obtained with the UPC-I trace using
the UPC-II trace for training. In particular, the figure presents the classification
accuracy removing the error introduced by the inversion of the features. The
results of this experiment confirm that this source of error explains the bulk of the
classification error, especially for low sampling rates (from p = 0.1 to 0.001). The
following sections analyze other sources of error that explain the rest of the error
observed under sampling.
3.4.2
Changes in the Flow Size Distribution
A second source of inaccuracy is the impact of sampling on the flow size distribution. When using Sampled NetFlow, short flows in terms of packets (i.e., mice)
can be easily missed by the sampling process, while long flows (i.e., elephants)
are usually sampled, even when using very low sampling rates. As a result, the
distribution of elephant and mice flows in a sampled trace is different from that
of an unsampled (complete) trace. As we will see, this impacts the classification
accuracy of sampled traffic when an unsampled trace is used for training. Next,
we present a theoretical analysis that shows that this distribution changes significantly, even when using high sampling rates.
Let Xp be a discrete random variable that denotes the original size of the
sampled flows when using a sampling rate of p. Please note that Xp represents
the original number of packets of the sampled flows, not the amount of sampled
packets. Let Y be a discrete random variable that denotes the original flow length
distribution, with probability mass function P [Y = y]. Then, the probability mass
42
CHAPTER 3. TRAFFIC CLASSIFICATION WITH NETFLOW
1
+
o
Overall accuracy
0.8
+
o
+
+
+
+
o
o
+
0.6
o
o
0.4
o
0.2
o
+
Estimated Features
Real Features
0.0
1
0.5
0.1
0.05
0.01
Sampling rate(p)
0.005
0.001
Figure 3.5: Overall accuracy when removing the error introduced by the inversion
of the features (UPC-I trace, using UPC-II for training)
function of Xp can be expressed as:
Px
Bp (i, x)P [Y = x]
P [Xp = x] = P∞ i=1
Pj
j=1
k=1 Bp (k, j)P [X = j]
(3.11)
where Bp (k, n) is the binomial probability, defined as Bp (k, n) = nk pk (1 − p)(n−k) .
Figure 3.6 validates our model (Eq. 3.11) against the empirical flow length
distribution of the detected flows with p = 0.1 in the UPC-II trace. In this case,
we used the probability mass function P [Y = y] obtained empirically from the
UPC-II trace (without sampling). The figure confirms that the model fits very
well the empirical distribution under sampling.
Figure 3.6 also confirms our hypothesis for our trace. In order to have a more
general result we approximate the original flow length with a Pareto distribution [75]. Thus, we can express the probability mass function of Y as:
P [Y = y] =
αLα x−α−1
Lα
1− H
(3.12)
Figure 3.7 plots the flow length distribution of the detected flows for different
sampling probabilities. The figure was obtained by combining Eq. 3.11 with Eq.
3.12. The parameters of the Pareto distribution were approximated using the
3.4. ANALYSIS OF THE SOURCES OF INACCURACY UNDER SAMPLING43
0.35
Analytical PMF
Empirical PMF
0.3
Probability
0.25
0.2
0.15
0.1
0.05
0
0
10
20
30
40
50
60
Flow Length (packets)
70
80
90
100
Figure 3.6: Validation of Eq. 3.11 against the empirical distribution of the original
flow length detected with p = 0.1 (UPC-II trace).
empirical probability mass function of Y (α = 3.7003, β = 0.3650, L = 1 and
H = 1 000).
The figure shows that the distribution of detected mice and elephant flows
changes significantly, even with very high sampling probabilities. For high sampling rates (e.g., p = 0.1), the amount of mice flows is remarkably larger than that
for low sampling rates (e.g., p = 0.001). However, the probability of detecting
elephant flows remains highly unaffected with respect to the sampling probability.
This result confirms that the flow length distribution of the traffic sampled during the classification phase is significantly different from that of the (unsampled)
training set. The use of such imbalanced datasets is known to be an important
source of inaccuracy of supervised learning methods [76].
Recall the previous experiment presented in Figure 3.5, where we analyzed the
impact of the inversion of the traffic features. When p = 1, the traffic classification
44
CHAPTER 3. TRAFFIC CLASSIFICATION WITH NETFLOW
0.35
p=1
p = 0.1
p = 0.01
p = 0.001
0.3
Probability
0.25
0.2
0.15
0.1
0.05
0
0
10
20
30
40
50
60
Flow Length (packets)
70
80
90
100
Figure 3.7: Flow length distribution of the detected flows when using several
sampling probabilities. The figure was obtained combining Eq. 3.11 with Eq. 3.12
method achieves an overall accuracy above 0.9, while under sampling (even when
the features are perfectly accurate), the accuracy drops below 0.8. This decrease
in accuracy can be mostly attributed to the different distribution of the flows
observed in the classification and training phases.
3.4.3
Flow Splitting
The last source of inaccuracy that we observed under sampling is flow splitting.
NetFlow implements several flow expiration mechanisms [77]. One of them is
the inactivity timeout, which expires the flows after a certain amount of time
of inactivity (TTL seconds). That is, if no packet is received for a given flow
during TTL seconds, NetFlow expires the flow and reports it. Under certain
circumstances, the packet inter-arrival time of an active flow may be larger than the
3.4. ANALYSIS OF THE SOURCES OF INACCURACY UNDER SAMPLING45
TTL, and this would result in several instances of the same flow reported, instead
of just one. Packet sampling aggravates this problem, since a set of consecutive
packets may be missed (not sampled), which may cause NetFlow to incorrectly
expire the flow while it is still active.
In this case, the flow is split and reported twice (or more times) to the traffic classification method. This situation affects the classification accuracy, since
the features of split flows only reflect partial information of the actual flow (e.g.,
estimated size and duration of split flows is shorter).
In this section, we aim to model the flow splitting probability under sampling
and discuss its impact on the traffic classification method. First, we focus on
an analytical model for the probability of splitting a flow, either in one or more
parts, as a function of the sampling rate p. For this purpose, we consider the
average inter-arrival time of the packets within a flow (iat), the inactivity timeout
of NetFlow (TTL) and its length in packets (n).
We base our model in the well-known consecutive-k-out-of-n:F [78] theory,
which accounts for the reliability of a system composed of an ordered sequence
of n components, such that the system fails if, and only if, at least k consecutive
components fail. In our case, the components are the packets of a flow, and the flow
is split if at least k consecutive packets are not sampled, where k = dT T L/iate.
From the consecutive-k-out-of-n:F theory, we know that the probability P of
having at least one run of k missed packets out of n can be expressed with the
following recursive formula:
P (n, q, k) = P (n − 1, q, k) + q k (1 − q)(1 − P (n − k − 1, q, k))
(3.13)
with the following base cases:
• if n < k then P = 0
• if n = k then P = q k
where q is the probability of missing a packet (q = 1 − p). Since a flow cannot
be split unless at least 2 packets have been sampled, we can express the split
probability Psplit (n, p) as:
Psplit (n, p) = P (n, 1 − p, dT T L/iate)[1 − (Bp (0, n) + Bp (1, n))]
where Bp (k, n) is again the binomial probability, defined as Bp (k, n) =
p)(n−k) .
(3.14)
n
k
k
p (1 −
46
CHAPTER 3. TRAFFIC CLASSIFICATION WITH NETFLOW
In order to validate our model, we plot in Figure 3.8 the actual number of
split flows in the UPC-II trace as a function of the sampling rate p, along with
the values obtained with our model using Eq. 3.14. The figure confirms that
our model predicts remarkably well the number of split flows for all the sampling
rates. It is also interesting to observe that the amount of split flows is not directly
proportional to the sampling rate and shows a peak at p = 0.1.
In order to understand this relationship, we must introduce the notion of sparse
flow [79]. According to our results, split flows typically have very long durations
(in the order of hundreds of seconds), but contain very few packets (an average
of 16 in our traces). Such sparse flows are more prone to be split [79]. However,
the probability of capturing more than one packet of a sparse flow is very small
for low sampling rates, which explains the decrease of split flows in Figure 3.8 for
sampling rates below 0.1.
18000
Empirical
Analytical
16000
14000
Number of splits
12000
10000
8000
6000
4000
2000
0
0.5
0.1
0.05
0.01
Sampling Rate (p)
0.005
0.001
Figure 3.8: Amount of split flows as a function of the sampling probability p. The
figure shows both the empirical results (UPC-II trace) and the analytical ones
(Eq. 3.14)
Although this source of error decreases the classification accuracy of those flows
3.5. DEALING WITH SAMPLED NETFLOW
47
that have suffered a split (e.g., the accuracy of split flows is 8% below the average
for the UPC-II trace with p = 0.5), its impact in the overall classification accuracy
in our datasets was low. This is explained by the fact that the percentage of split
flows in our traces is small (about 1% with p = 0.5), since they contain very few
sparse flows. However, flow splitting can be an important source of inaccuracy in
other traffic profiles with a larger portion of sparse flows.
3.5
Dealing with Sampled NetFlow
Given the severe impact of packet sampling on the classification accuracy observed
in previous sections, this section presents a simple improvement in the training
process that significantly increases the performance of our traffic classification
method in the presence of sampling (e.g., Sampled NetFlow). Finally, we also
review some of the limitations detected during the evaluation of our method, which
are common to most ML-based techniques.
3.5.1
Improving the Classification Method
Unlike in previous works, where a set of complete (unsampled) flows are used in
the training phase (e.g., [15]), we propose instead to apply packet sampling to the
training process. This simple solution is possible because Sampled NetFlow uses a
static sampling rate [54], which must be set at configuration time by the network
operator. Therefore, we can use exactly the same sampling rate in both the training
and classification processes. This solution has the advantage of circumventing the
sources of inaccuracy discussed in Section 3.4, since in this case the features do
not need to be inverted, while the same flow distribution and split probability is
maintained in both the training and classification sets
In particular, we repeat the training process described in Section 3.2 by applying a set of sampling rates commonly used by network operators, which range from
50% to 0.1%, and then evaluate the accuracy of the classification process under
these static sampling rates.
Figure 3.9 shows the substantial improvement obtained in terms of classified
flows when applying sampling to the training set. Remarkably, we can observe
that the flow accuracy degrades much more gracefully compared to the original
technique. For example, with p = 0.01, the accuracy per flow increases from
46.96% to 82.38%. Similar improvements are obtained for the other sampling
rates. According to the theoretical analysis presented to Section 3.4, this increase
48
CHAPTER 3. TRAFFIC CLASSIFICATION WITH NETFLOW
Figure 3.9: Overall accuracy (mean with 95% CI) of our traffic classification
method with a normal and sampled training set
Figure 3.10: Precision (mean with 95% CI) by application group (per flow) of our
traffic classification method with a sampled training set
of accuracy can be mostly attributed to the error introduced by the inversion
of the traffic features when using an unsampled training set. The enhancement
is also partially explained by the already mentioned elephant-mice phenomenon,
since now the proportion of elephant and mice flows is kept similar in the training
and classification phases. Although the results in terms of packets and bytes are
more modest, especially for p = 0.01, the improvement is still considerable. This
lower gain is expected since the accuracy in terms of packets and bytes under
aggressive sampling rates is mainly driven by large flows, which are less affected
by our enhancement in the training phase.
The precision (Figure 3.10) and recall (Figure 3.11) are also significantly higher
for all application groups and sampling rates. Note also that, while our initial
3.5. DEALING WITH SAMPLED NETFLOW
49
Figure 3.11: Recall (mean with 95% CI) by application group (per flow) of our
traffic classification method with a sampled training set
classification method was not able to identify almost any flow of some application groups (e.g., Games, Chat or Others) in the presence of sampling (see Figures 3.2 and 3.3), with the proposed improvement in the training process the precision and recall of these applications are notably improved (e.g., precision from
2.78% to 83.97% for Games applications with p = 0.01).
3.5.2
Evaluation in Other Network Environments
In previous sections, we showed the performance of our traffic classification method
using different traces from a single network viewpoint (UPC). The objective of
this section is to evaluate if similar results are also obtained in other network
environments.
For this purpose, we used 5 packet-level traces from two public repositories
(CAIDA [69] and WITS [70]). The traces from the CAIDA archive were collected in
two OC-192 backbone links from a Tier-1 ISP in the US. Conversely, the AucklandVIII trace from the WITS repository was collected in a 100 Mbps access link of the
University of Auckland to the global Internet. Details of the traces are presented
in Table 3.6.
A common problem when using public traces is that packet payloads are usually
not available for privacy reasons. However, packet payloads are needed to build
the ground-truth used to measure the classification accuracy. In order to address
this limitation, we labeled the traces in Table 3.6 using our own classifier, which
we trained using the UPC-II trace as described in Section 3.2.5. Although this
approach assumes that the accuracy is 1 when p = 1, it is sufficient to analyze the
decrease of accuracy due to packet sampling.
Figure 3.12 presents the overall accuracy, in terms of classified flows, when us-
50
CHAPTER 3. TRAFFIC CLASSIFICATION WITH NETFLOW
Table 3.6: Characteristics of the traces from other environments
Name
CAIDA Chicago 08
CAIDA Chicago 09
CAIDA SanJose 08
CAIDA SanJose 09
Auckland-VIII
Date
18-12-08
16-07-09
21-08-08
17-09-09
10-12-03
Day
Thu
Thu
Thu
Thu
Wed
Start Time
14:00
15:00
15:00
15:00
15:00
Duration
1h
1h
1h
1h
4h
Packets
3.53 G
4.28 G
2.32 G
3.67 G
35 M
Bytes
2.7 T
4.2 T
1.3 T
2.4 T
16 G
Avg. Util
6.1 Gbps
9.3 Gbps
2.9 Gbps
5.4 Gbps
9.2 Mbps
ing a normal (unsampled) and a sampled training set with the traces in Table 3.6.
The figure shows that similar improvements to those observed in Figure 3.9 with
the traces from UPC are also obtained with the traces from other network scenarios. In general, the accuracy when using a complete training set decreases abruptly
for moderate sampling rates and starts to recover slowly for low sampling rates.
This effect, which was already observed with the traces from UPC in Figure 3.4, is
explained by the fact that, as discussed in Section 3.4, the error in the inversion of
the flow features is much smaller for elephants flows, which are the most predominant with low sampling rates (e.g., 1/1 000). In contrast, the accuracy when using
the improvement presented in Section 3.5.1 is more stable and is kept reasonably
high for all sampling rates.
3.5.3
Lessons Learned and Limitations
In the course of this study, we have found that several limitations of most ML
techniques for traffic classification are directly related to the ground-truth used in
the training phase. This explains the accuracy problems observed in Sections 3.3
and 3.5 for some application groups and metrics. As a result of such experience, we
present some simple solutions that could help to improve the quality of the training
set in future works: (i) include a similar number of instances of all application
groups, (ii) increase the number of instances of elephant flows, for example by
using longer traces that can better capture the behavior of large flows, which have
a high probability of being ignored in small traces due to the sanitization process,
(iii) mix several instances of multiple traces from different hours, days and network
scenarios in the training set, (iv) define alternative groups than those used by L7filter, according to the actual behavior of the applications, and avoid generic groups
(e.g., Others), and (v) investigate better, but still automatic, labeling techniques
3.5. DEALING WITH SAMPLED NETFLOW
0.8
1
o+
+
+
+
+
0.6
o
0.4
o
+
0.2
0.0
0.5
0.1
o
Overall accuracy
Overall accuracy
1
51
o
o
+
o
Normal CAIDA_Chicago_08
Sampled CAIDA_Chicago_08
0.05
0.01
0.005
0.8
+
+
+
o
o
o
+
0.6
0.4
o
+
0.2
0.0
0.001
+
o
0.5
0.1
Sampling rate(p)
o+
+
o
+
+
o
o
+
o
Overall accuracy
Overall accuracy
Normal CAIDA_Chicago_09
Sampled CAIDA_Chicago_09
0.05
0.01
0.005
0.001
1
0.6
o+
0.4
o
+
0.2
0.0
o+
Sampling rate(p)
1
0.8
o
0.5
0.1
Normal CAIDA_Sanjose_08
Sampled CAIDA_Sanjose_08
0.05
0.01
0.005
0.8
+
+
+
o
o
o
+
0.6
o
+
o
0.4
o
+
0.2
0.0
0.001
o+
0.5
0.1
Sampling rate(p)
Normal CAIDA_Sanjose_09
Sampled CAIDA_Sanjose_09
0.05
0.01
0.005
0.001
Sampling rate(p)
1
Overall accuracy
+
0.8
+
+
o
+
+
o
o
0.6
o
+
o
o
0.4
o
+
0.2
0.0
0.5
0.1
0.05
Normal Auckland−VIII
Sampled Auckland−VIII
0.01
0.005
0.001
Sampling rate(p)
Figure 3.12: Overall accuracy with a normal and sampled training set using the
traces from other environments
for the training phase (e.g., [9, 67]).1
1
Note that the accuracy of supervised learning methods is directly restricted by the labeling
technique used in the training phase. For example, if DPI techniques are used for training, it
is improbable that the resulting classification method can identify encrypted traffic, unless it
52
3.6
CHAPTER 3. TRAFFIC CLASSIFICATION WITH NETFLOW
Related Work
To the best of our knowledge, only a workshop paper [15] has analyzed the traffic
classification problem in the presence of sampling. In particular, [15] evaluated the
impact of packet sampling on the Naı̈ve Bayes Kernel Estimation supervised ML
technique. In this work, we analyze instead the impact of sampling on the C4.5
classification method, which is faster in terms of both training and classification
time. Despite the similarities between both techniques, we have observed a significantly higher impact of sampling in the classification accuracy. While in [15] a
complete, unsampled trace is used in the training phase, in this work we show that
this approach has a severe impact on the classification accuracy in the presence of
sampling. Since the trace used in [15] is not publicly available to further analyze
these differences, we attribute them to the more complete datasets used in our
study, which are more difficult to classify using only the port numbers (note that
port numbers are resilient to sampling) and include a much larger volume of traffic
and both TCP and UDP connections. In contrast, [15] only used a single trace
that did not contain UDP traffic. In order to allow for further comparison of our
results with other classification techniques, we have made our datasets publicly
available to the research community. They can be obtained at [68].
The impact of sampling has also been studied in other, closely related research
areas. In the field of anomaly detection, several works have analyzed the impact
of different sampling techniques on the performance of portscan detection [80–82].
They also conclude that packet sampling has an important impact on the detection
accuracy, increasing both false negative and false positive ratios. Although some
studies reported lower impacts when using flow sampling (e.g., [80, 81]), Sampled
NetFlow only supports packet sampling (systematic and random) [77]. Androulidakis et al. [83] show that systematic sampling is especially problematic when the
detection algorithms depend on the observation of a particular packet (e.g., SYN
flag). Brauckhoff et al. [84] also find that some anomaly detection metrics are
more resilient to sampling than others, especially those based on entropy summarizations, and that detection algorithms based on packet and byte counts are less
affected than those based on flow counts.
Finally, in the field of bandwidth estimation, Davy et al. [85] explored the
problem of estimating the bandwidth demand from sampled network accounting
data for QoS-aware network planning. Their results show estimation errors in the
±10% range. They also find that the estimation error depends on the particular
exhibits similar behavior than the unencrypted traffic of the same application.
3.7. CHAPTER SUMMARY
53
class of traffic and that it is larger for short flows at low sampling rates.
3.7
Chapter Summary
In this chapter, we addressed the traffic classification problem with NetFlow data
using a well-known supervised learning technique. Our results allow us to come to
the conclusion that: (i) supervised methods (e.g., C4.5) can achieve high accuracy
(≈90%) with unsampled NetFlow data, despite the limited information provided
by NetFlow, as compared to the packet-level data used in previous studies, and (ii)
the impact of packet sampling on the classification accuracy of supervised learning
methods is severe.
Therefore, we proposed a simple improvement in the training process that significantly increased the accuracy of our classification method under packet sampling. For example, for a sampling rate of 1/100, we achieved an overall accuracy
of 85% in terms of classified flows, while before we could not reach an accuracy
beyond 50%.
Given that Sampled NetFlow is the most widely extended monitoring solution
among network operators, we hope that these encouraging results can incentivize
the deployment of more accurate techniques in operational networks, where the
obsolete port-based method is often the unique real solution.
54
CHAPTER 3. TRAFFIC CLASSIFICATION WITH NETFLOW
Chapter 4
Autonomic Traffic Classification
System
4.1
Introduction
Once addressed the deployment problem in Chapter 3, in this chapter, we address the traffic classification problem with special emphasis on the maintenance
of the technique. We propose a complete realistic solution for network operation
and management that can be easily deployed and maintained. The classification method combines the advantages of multiple methods, while minimizing their
limitations. As described previously in Chapter 3, we use NetFlow as input for
the classification to facilitate the deployment. In addition, we propose an autonomic retraining system that can sustain a high classification accuracy during
long periods of time. To the best of our knowledge, this is the first solution for
traffic classification to provide this feature, which is central for network operation
and management, because it removes the need of human intervention of previous solutions, and makes the system easier to maintain. Finally, we evaluate the
performance of our method using large traffic traces from a production network.
The rest of this chapter is organized as follows. The proposed classification
technique and the autonomic retraining system are described in Section 4.2. Section 4.3 presents a performance evaluation of our method and analyzes the impact
of different retraining policies on its accuracy. Section 4.4 reviews in greater detail
the related work. Finally, Section 4.5 concludes the chapter.
55
56
CHAPTER 4. AUTONOMIC TRAFFIC CLASSIFICATION SYSTEM
Classification
path
APPLICATION
IDENTIFIER
MONITORING TOOL
NETFLOW v5
DATA
Training
path
FLOW SAMPLING
(FULL PACKET)
PACKETS
FLOW FEATURES
EXTRACTION
FLOW
STATISTICS
TRAINING DATA
RETRAINING
MANAGER
PACKETS
AUTONOMIC
RETRAINING
SYSTEM
FLOW
LABELLING
LABELED
FLOWS
A. I. LIB
UPDATE
STORE
LABELED
FLOWS
ACTIVATE
TRAINING
CLASSIFY FLOWS
APPLICATION
IDENTIFIER
TRAINER
NEW
TRAINED
MODELS
A. I. LIB
BUILDER
A. I. LIB
UPDATE
Figure 4.1: Application Identifier and Autonomic Retraining System architecture
4.2
Traffic Classification System
This section describes our traffic classification method for Sampled NetFlow and
its autonomic retraining system. Figure 4.1 illustrates the architecture of the
complete traffic classification system, which is divided into two different parts.
Above in Fig. 4.1, the classification path is in charge of classifying the traffic online.
In order to achieve this goal we implement a classifier module, called Application
Identifier. This module (described later in Section 4.2.1) is loaded as a dynamic
library in the monitoring tool and it is only fed with NetFlow v5 data. Given that
it only needs the information provided by NetFlow, either sampled or not, and
does not have to track the flows, the system is swift and lightweight enough to be
deployed in large production networks. As a use-case, we integrated this module
in the SMARTxAC system [73] and evaluated it with data from a large production
university network, as we describe later in Section 4.3.
Below in Fig. 4.1, the training path carries flow-sampled raw traffic to the Autonomic Retraining System. This element has two important goals. First, it will
provide online information about the current accuracy of the Application Identifier
that is classifying the traffic online in parallel, as will be described later in Section 4.2.2. Second, it will generate a new classification model when this accuracy
falls below a threshold, as we describe in Section 4.2.2. Given that the Application
Identifier is loaded dynamically it could be reloaded at any time. This capability
along with the Autonomic Retraining System makes our system resilient to traffic
variations as, when the accuracy falls, the system will be automatically retrained
4.2. TRAFFIC CLASSIFICATION SYSTEM
57
and the classification model updated.
4.2.1
The Application Identifier
The Application Identifier is the module in charge of the online classification. As
aforementioned, it combines different techniques for the sake of exploiting their
advantages and reducing their practical limitations. For the reason discussed before, we only consider use methods capable of dealing with Sampled NetFlow data.
Furthermore, we need fast classification techniques and methods with a lightweight
training phase. Although these constraints complicate the classification, this makes
the system easier to deploy and maintain, which are crucial features for network
operation and management. The final choice consists of the combination of three
techniques of different nature with some improvements especially made to increase
their accuracy with Sampled NetFlow data.
Firstly, we use an IP-based technique based on the proposal presented by Mori
et al. in [41]. Basically, this technique tracks down in an offline phase the IP
addresses belonging to famous web sites (e.g., Google, Megavideo). This technique
is very accurate, however its completeness has been significantly degraded given the
migration of some applications to Content Delivery Networks. This has relegated
its use in our system as a technique to be combined with other more complex and
accurate techniques.
The second method used by the Application Identifier is an adaptation of the
Service-based classifier described by Yoon et al. in [42]. A service is defined
as the triplet <IP, Port, Protocol> assigned to a specific application. The list
of services is also created in an offline phase using a dataset of labeled flows as
follows. In our system, we aggregate all the available flows by their triplet and
then, a service is created when a triplet has a minimum number of flows (n)
and there is a predominant label (> m%). Unlike in [42], we do not use a time
threshold but we require a higher number of flows (n ≥ 5). We studied offline an
efficient configuration of those parameters in order to increase the completeness
of the technique while keeping high accuracy. The results have determined that a
proper configuration in our setting is selecting n = 10 and m > 95%.
The last method used in the Application Identifier is a ML technique, namely,
the C5.0 decision tree, an optimized successor of the well-known C4.5 presented
by Quinlan in [71]. To the best of our knowledge, this is the first work in the field
of traffic classification that uses this variant. Nevertheless, several papers have
previously highlighted the advantages of its predecessor. Kim et al. in [36] and
Williams et al. in [32] compared different classification techniques showing that
58
CHAPTER 4. AUTONOMIC TRAFFIC CLASSIFICATION SYSTEM
C4.5 achieves very high accuracy and classification speed due to its inherent feature discretization [86]. Furthermore, the C5.0 is characterized by having shorter
training times compared with its predecessor [87] and with other well-known ML
techniques as Support Vectors Machines or Neural Networks [36]. This ability
is a key point in our proposal given its importance in the Autonomic Retraining
System. Because of this, we have not applied in our evaluation any improving
technique as boosting or bagging. The accuracy obtained by the C5.0 with the
default configuration is already very high and the improvement obtained by these
techniques was negligible compared with the critical increase of training time. Another important feature of our ML-based technique is its full completeness. As
mentioned above, the IP and Service-based techniques have limited completeness
given that they rely on IP addresses. However, the ML-based technique allows the
Application Identifier to classify all the traffic.
It is important to recall that a requirement of our system is that the Application Identifier has to work only with Sampled NetFlow traffic. The IP and
Service-based techniques work properly with Sampled NetFlow because the triplet
<IP, Port, Protocol> is not affected by the sampling. However, the ML technique is substantially affected by this constraint, given that NetFlow v5 reduces
the amount of features available for the classification and applying sampling considerably impacts on the computation of the features [15, 46]. In order to address
this limitation, we have implemented the C5.0 ML technique following the recommendations proposed in [46] to improve the classification under Sampled NetFlow,
which basically consists of applying sampling to the training phase.
In order to combine the power of the three techniques, we combine their classification in a final decision. We give priority to the IP-based technique given the
IP addresses have been manually checked. However, given its low completeness,
most of the traffic is classified by the Service and ML-based techniques. The distribution of the traffic classified by each technique changes with each retraining,
however their contributions are usually around 10% for the IP-based, 60% for the
Service-based, and 30% for the ML-based technique. Those techniques give their
classification decision with a confidence value. The classification decision with
highest confidence is selected.
Table 4.1 presents the features used by each technique, all of them obtained
from NetFlow v5 data. This is very important because it allows the Application
Identifier to be very lightweight and easy to deploy given that it works at flow-level
and does not have to compute the features for the classification.
4.2. TRAFFIC CLASSIFICATION SYSTEM
59
Table 4.1: Features used by each classification technique in the Autonomic Traffic
Classification System
Technique
IP-based
Service-based
ML-based
4.2.2
Features
IP addresses
IP, Port and Protocol
NetFlow v5 features +
average packet size + flow time
+ flow rate + inter-arrival time
The Autonomic Retraining System
A common limitation of previous proposals presented in the literature, including
the ML techniques proposed in [46] in which our Application Identifier is based on,
is that they usually require an expensive training phase, which involves manual
inspection of a potentially very large number of connections. In order to automate
this training phase, we developed a retraining system that does not rely on human
supervision. This property together with the ability to classify Sampled NetFlow
data, makes our proposal a realistic solution for network operators and managers.
Unlike the Application Identifier, the Autonomic Retraining System presented
in Fig. 4.1 uses packet-level data as input. This data is labeled with DPI-based
techniques and later used to build the ground-truth for future retrainings. Applying those techniques online is unfeasible given the high resource consumption.
However, our system is able to retrain itself and sustain high accuracy rates along
time with very few data. This allows us to apply an aggressive flow sampling rate
to the Autonomic Retraining System input keeping the system very lightweight
and economically feasible for the operation and management of large production
networks.
The Autonomic Retraining System is divided in three phases. The first one
corresponds to the labeling and feature extraction, the second checks the accuracy
and stores the ground-truth data and, finally, the last phase retrains and reloads
the classifier when it is necessary.
In the labeling and feature extraction phases, the input data is processed in two
different ways. On the one hand, while aggregating the data per flow, a feature
extraction is applied to obtain the NetFlow v5 features that would be obtained in
the classification path. On the other hand, to obtain a reliable ground-truth, we
use a set of DPI techniques, including PACE, a commercial DPI library provided
by ipoque, which is known to have high accuracy with low false positive ratio [10].
60
CHAPTER 4. AUTONOMIC TRAFFIC CLASSIFICATION SYSTEM
Table 4.2: Application groups and traffic mix in the UPC-II and CESCA datasets
Group
Applications
Web
DD
Multimedia
P2P
Mail
Bulk
VoIP
DNS
chat
Games
Encryption
Others
HTTP
E.g., Megaupload, MediaFire
E.g., Flash, Spotify, Sopcast
E.g., Bittorrent, eDonkey
E.g., IMAP, POP3
E.g., FTP, AFTP
E.g., Skype, Viber
DNS
E.g., Jabber, MSN Messenger
E.g., Steam, WoW
E.g., SSL, OpenVPN
E.g., Citrix, VNC
Flows
UPC-II CESCA
678 863 17 198 845
2 168
40 239
20 228
1 126 742
877 383 4 851 103
19 829
753 075
1 798
27 265
411 083
3 385 206
287 437 15 863 799
12 304
196 731
2 880
14 437
71 491
3 440 667
55 829
2 437 664
Moreover, to increase the completeness of the DPI classification, we added two
extra libraries, OpenDPI [9] and L7-filter [8]. In addition, our system is extensible
and allows the addition of new labeling techniques to increase the completeness
and accuracy. Based on their relative accuracy, we have given the highest priority
to PACE and the lowest priority to L7-filter. The final label of each flow is selected
from the DPI technique with highest priority. An evaluation of the impact of the
different DPI techniques used in the Autonomic Retraining System is presented in
Section 4.3.1.
In the second phase, the retraining manager (see Fig. 4.1) receives the labeled
flows together with their NetFlow v5 features. Those that are not labeled as unknown are stored for future retrainings. In parallel, the retraining manager sends
the flows together with their NetFlow v5 features to the Application Identifier.
This Application Identifier is identical to the one that is currently running in the
monitoring tool. The Application Identifier classifies the flow by obtaining a second label. This label is the same label that the monitoring tool would obtain.
By comparing both labels, we can compute the actual accuracy of the system.
When the accuracy falls under a threshold, we create a new trainer in order to
build a new classification model. The accuracy in our system is computed from
the last flows seen (e.g., 50K in our evaluation). Although the classification is
done at the application level, the accuracy is computed aggregating the results at
4.2. TRAFFIC CLASSIFICATION SYSTEM
61
the group level as described in Table 4.2. Table 4.2 also presents the traffic mix
of the traces used in the evaluation. These traces that are further described in
Section 4.3, although collected in a research/university network, are compounded
by a heterogeneous mixture of applications.
A trainer runs as a separate thread and, using the ground-truth dumped in
the previous phase, retrains the ML-based and Service-based techniques. The
generation of the training dataset is a key point of the retraining system given the
important impact it has on the perdurability and accuracy of the models. This
process is described in detail in Section 4.2.3. Once the new classification models
are built, the new classification library is compiled and dynamically loaded in the
Application Identifiers that are running on the monitoring tool and the Autonomic
Retraining System itself. An evaluation of the cost of this process is presented in
Section 4.3.2.
4.2.3
Training Dataset Generation
The proper selection of the instances that compose the training dataset will considerably impact on the quality and perdurability of the new classification models
created. This way, we have studied two features to build the training dataset: the
retraining policy and the training size of the dataset.
The training size is the number of instances (i.e., labeled flows together with
their NetFlow v5 features) that compose the training dataset. We refer to the
training size as X. The training size substantially impacts on the training times
and the quality of the models. On the one hand, selecting a small training size
would produce a system highly reactive to accuracy falls given that the retraining
time is shorter. However, the classification models built would be less accurate as
they have less information (i.e., instances) to build it. On the other hand, a bigger
training size would increase the training times but produce more accurate models.
Regarding the retraining policy, we implemented two different policies to perform the retraining. The first approach takes into account the last X labeled flows.
However, this approach could be biased to the last traffic received. We refer to it
as the naive retraining policy. The second approach uses random flows from the
last Y days, as follows called long-term retraining policy. Although it is totally
configurable, for the sake of a fair comparison, we also select a total of X flows
proportionally distributed in Y days, where Xi is the number of flows for the day
i:
2(Y −i)
Xi = X · Y
2 −1
62
CHAPTER 4. AUTONOMIC TRAFFIC CLASSIFICATION SYSTEM
Thus, it creates a training data set in which recent days have more weight (i.e.,
more instances) than older ones.
Section 4.3.2 evaluates the impact of those features presenting sound conclusions about the best trade-off between accuracy and performance to obtain proper
datasets to maintain an accurate online traffic classifier for a network management
tool.
4.3
Evaluation
In this section, we first evaluate the contribution and impact of the different DPI
techniques used in the Autonomic Retraining System. Then, we evaluate the impact of the policies presented in Section 4.2.3 on the generation of the training
dataset. The obtained results are then used to select a proper configuration for
the training dataset. Afterwards, we analyze the impact of the Autonomic Retraining System on the Application Identifier. This evaluation is performed for
both sampled and unsampled scenarios. The results show the effectiveness of our
system as an autonomous and accurate traffic classifier for large networks.
4.3.1
Evaluation of Labeling DPI-based Techniques
DPI-based techniques are scarcely used for online classification given their high
resources consumption. However, these techniques are commonly used as an automatic ground-truth generator [36,46,64]. In our system, the Autonomic Retraining
System uses three DPI-based techniques (i.e., PACE, OpenDPI, and L7-filter) to
generate the ground-truth online. This is feasible given that the Autonomic Retraining System only needs a small sample of the traffic to maintain the Application
Identifier updated. Samples are selected by applying a high flow sampling rate
to the training path. This extremely reduces the amount of traffic to be analyzed
compared with the whole traffic received by the classification path. This type of
sampling preserves the entire payload of the flows, allowing DPI-based techniques
work properly.
The experiments in this section use the trace CESCA. The CESCA trace,
described in detail in Section 2.3.2, is a fourteen-days packet trace collected on
February 2011 in the 10-Gigabit access link of the Anella Cientı́fica, which connects
the Catalan Research and Education Network with the Spanish Research and
Education Network. As described in Section 4.2.2, the Autonomic Retraining
System only requires a small sample of the traffic to achieve its goal. For this
4.3. EVALUATION
63
PACE
L7_FILTER
17.25%
PACE
L7_FILTER
4.41%
2.24%
49.23%
12.16%
3.46%
26.65%
2.24%
0.04%
52.58%
29.34%
0.04%
0.18%
OPENDPI
(a) By application
0.18%
OPENDPI
(b) By group of application
Figure 4.2: DPI labeling contribution in the CESCA dataset
reason, similarly to the flow sampling applied to the training path, we applied a
1/400 flow sampling rate. Although the Autonomic Retraining System can handle
higher flow sampling rates, we applied this one because it was the lowest that
allowed us to collect the trace without packet loss in our hardware.
Figure 4.2a and 4.2b shows the contribution of the different DPI techniques
in the ground-truth generation. The major contributor in the labeling process is
PACE. As Fig. 4.2a shows, the contribution of OpenDPI and L7-filter are very low,
0.22% (0.18% + 0.04%) and 2.24% respectively. This is because most of the labels
of these techniques match with the labels of PACE and PACE has higher priority
(Sec. 4.2.2). Figure 4.2b shows that OpenDPI and L7-filter miss some application
labels but match them at group level. This can be seen in the decrease of the
PACE percentage. These results also help us to understand the completeness our
system would achieve in case we have no access to a commercial labeling tool.
In order to guide the network operator in the selection of an appropriate flow
sampling rate for their network, Table 4.3 presents the consumption of the DPI
techniques by profiling the Autonomic Retraining System running in a 3GHz machine with 4GB of RAM. Table 4.3 shows that the average consumption of the
different DPI techniques has the same order of magnitude. However, looking at
the standard deviation and the maximum (Max) by flow, L7-filter behaves totally
different than PACE and OpenDPI. This is because L7-filter has been limited
to the first packets and bytes of each flow in order to reduce the false positive
64
CHAPTER 4. AUTONOMIC TRAFFIC CLASSIFICATION SYSTEM
Table 4.3: DPI techniques consumption in the Autonomic Retraining System
Metric
Flow
Packet
Avg.(µs/flow)
STD(µs/flow)
Max (µs)
Avg.(µs/packet)
STD(µs/packet)
Max (µs)
DPI techniques
L7-filter OpenDPI
PACE
34.54
25.92
32.36
41.29
1 419.10
1 721.86
13 118
1 369 695 1 558 510
1.74
1.29
1.66
10.49
4.31
4.87
13 118
13 168
12 979
ratio [46]. On the other hand, OpenDPI and PACE perform a more thorough examination in order to find out the application label. In more restrictive scenarios,
OpenDPI and L7-filter could be deactivated to improve the performance of the
system. However, given that OpenDPI and L7-Filter detects some applications
that PACE does not, we have included both DPI techniques in the system. For
instance, the 14-days CESCA trace contains 71 million flows (with 1/400 flow sampling applied). Without sampling, 42 µs would be needed in average to process
each flow without packet loss (14 days / (71 million flows x 400) = 42 µs/flow).
Table 4.3 shows that only the DPI libraries require 92 µs per flow in average.
This shows that a traffic classification system based solely on DPI would not be
sustainable in our network scenario and it does not scale well to higher link speeds.
4.3.2
Training Dataset Evaluation
In this section, we evaluate the impact of the policies presented in Sec. 4.2.3 in
the generation of the training dataset used by the Autonomic Retraining System.
In all experiments, we use the trace UPC-II for the initial offline training and the
trace CESCA for the evaluation. As described in Section 2.3.1, the trace UPC-II
is a fifteen-minutes full-payload trace with more than 3 millions of flows. We used
different traces for the training and the evaluation in order to show the ability of
our system to automatically adapt itself to new scenarios.
In order to asses the quality of the system, the Autonomic Retraining System
uses the accuracy metric. As already mentioned in Sec. 4.2.2, the Autonomic
Retraining System computes the accuracy by calculating the number of correctly
classified flows from the last flows seen (i.e., 50K in our evaluation). The exact
definition of the accuracy metric is described in Section 2.2.1.
4.3. EVALUATION
65
Although the accuracy is the most popular metric used in the network traffic
classification literature it has some limitations. In order to confirm the quality of
the Autonomic Retraining System, we also compute the Kappa coefficient. This
metric is considered to be more robust because it takes into account the correct
classifications occurring by chance. The computation of the Kappa coefficient is
also described in Section 2.2.1.
The Kappa coefficient takes values close to 0 if the classification is mainly due to
chance agreement. On the other hand, if the classification is due the discriminative
power of the classification technique then the values are close to 1.
In order to evaluate the impact of the different retraining policies on the Autonomic Retraining System, we have performed a study of the impact of the parameter Y in the long-term retraining policy. The study evaluates the performance of
different values of Y (i.e., 5, 7, 9, 11) with a fixed accuracy threshold (i.e., 98%)
and a fixed training size (i.e.,X =500K). The results of this evaluation, presented
in Table 4.4, show that this parameter has not critic impact on the Autonomic
Retraining System. However, the values Y = 7 and Y = 11 achieve the highest
accuracies, being Y = 7 faster in the training process. As a result, we selected Y
= 7 for the long-term retraining policy. This way, the retraining is performed with
flows processed during the last seven days, allowing the system to cover the traffic
of an entire week.
Table 4.5 presents the results of the evaluation using three different training
sizes (i.e., X ={100K, 500K, 1M}) and two retraining policies (i.e., naive retraining
policy and long-term retraining policy). The evaluation has been performed using a
high retraining threshold (i.e., the Application Identifier is retrained if the accuracy
goes below 98%) in order to stress the system to perform multiple retrainings
by highlighting the differences between the different configurations. Unlike we
initially expected, Table 4.5 shows that the long-term retraining policy performs
slightly worse than the naive retraining policy in terms of accuracy. Moreover,
the average training time is shorter for the naive retraining policy. This is mainly
due the creation of the dataset that, although it could be optimized for the longterm retraining policy, it will be always longer than the naive retraining policy.
Regarding the training sizes, the option X=100K achieves lower average accuracy
than the other training sizes. However, the X=100K training size obtains the
highest minimum accuracy and the lowest average training time. This could be
interesting if the network demands a fast-recovery system to an accuracy fall.
The results comparing X=500K, 1M, and the naive retraining policy show that
these configurations obtain similar average accuracy. However, we have decided to
choose X=500K and naive retraining policy as the optimum configuration given
66
CHAPTER 4. AUTONOMIC TRAFFIC CLASSIFICATION SYSTEM
Table 4.4: Long-Term policy evaluation in the Autonomic Traffic Classification
System
Metric
Avg. Accuracy
Min. Accuracy
# Retrainings
Avg. Training Time
Cohen’s Kappa (k)
5 days
98.04%
95.64%
126
229 s
0.9635
Training Policy
7 days 9 days 11 days
98.12% 98.07% 98.12%
95.44% 95.44% 95.42%
125
125
125
232 s
234 s
242 s
0.9634 0.9634
0.9633
that it offers slightly better results. Regarding the impact of this policy on the
system, the training with a 98% accuracy threshold only requires 3.93h compared
to the 336h (14 days) of duration of the whole experiment, which represents only
13% of the total trace time. If the threshold is lowered up to 96%, the training
time is reduced to 0.54h (1.8% of the total trace time).
Although the Autonomic Retraining System bases its decisions on the accuracy metric, we have also computed the Kappa coefficient. Table 4.5 shows that
the values of the Kappa coefficient are very close to 1. This result confirms the
actual classification power of the Autonomic Retraining System showing that its
classification is not just due to chance agreement.
4.3.3
Retraining Evaluation
So far, we have separately studied the performance of the labeling techniques and
the impact of the different policies on the training dataset generation. Based on
these results, we selected a final configuration: the Autonomic Retraining System
uses the three DPI-based techniques (i.e., PACE, OpenDPI and L7-filter) for the
labeling process (Sec. 4.3.1), 500K flows as training size (i.e., X = 500K), and the
naive retraining policy (Sec. 4.3.2). In this section, we evaluate the Application
Identifier and the impact of the Autonomic Retraining System on its accuracy
with both sampled and unsampled traffic.
As discussed in Sec. 4.3.2, we use in all experiments the trace UPC-II for the
initial training and the traceCESCA for the evaluation. It is important to note
that the trace UPC-II was collected in December 2008 while the trace CESCA
was collected in February 2011. As a result, the system usually performs an initial
retraining to update the initial outdated model. This decision has been taken in
4.3. EVALUATION
67
Table 4.5: Training dataset evaluation in the Autonomic Traffic Classification
System
Training
Size
100K
500K
1M
Metric
Avg. Accuracy
Min. Accuracy
# Retrainings
Avg. Training Time
Cohen’s Kappa (k)
Avg. Accuracy
Min. Accuracy
# Retrainings
Avg. Training Time
Cohen’s Kappa (k)
Avg. Accuracy
Min. Accuracy
# Retrainings
Avg. Training Time
Cohen’s Kappa (k)
Training Policy
Long-Term Policy Naive Policy
97.57%
98.00%
95.95%
97.01%
688
525
88 s
25 s
0.9622
0.9567
98.12%
98.26%
95.44%
95.70%
125
108
232 s
131 s
0.9634
0.9652
98.18%
98.26%
94.78%
94.89%
61
67
485 s
262 s
0.9640
0.9650
order to show the impact of the spatial obsolescence, showing that in order to
obtain the most accurate classification model it is crucial to train the system with
the traffic of the scenario that it is going to be monitored.
Figure 4.3a presents the evaluation of the Application Identifier when no packet
sampling is applied to the traffic. We tested different accuracy thresholds in order
to show the behavior of the system depending on the preferences of the network
operator. The system maintains the accuracy of 94% by performing five retrainings
during the 14 days. With the 96% threshold it is able to sustain the accuracy
during long periods of time with only 15 retrainings. Using the highest threshold,
our method achieves better average accuracy than previous thresholds. However,
it is not capable to continuously maintain the 98% accuracy. Because of this,
the Autonomic Retraining System is almost continuously updating the classifier.
Nevertheless, these continuous retrainings have not any impact on the Application
Identifier giving that this procedure is done completely apart. Figure 4.3a also
shows the effectiveness of the retrainings, pointed out with cross symbols, that
usually produce a substantial increment of accuracy. An interesting result seen
68
CHAPTER 4. AUTONOMIC TRAFFIC CLASSIFICATION SYSTEM
by the 94% threshold is the ability of the system to automatically find a proper
classification model. As can be seen at the left part of the Fig. 4.3a, the system
performs three retrainings but finally builds a model that remains stable for about
a week.
We have also evaluated the performance of our system when packet sampling
is applied in the classification path. We perform the experiments with a common
sampling rate of 1/1 000 using the configuration before described. Figure 4.3b
shows the impact of the retraining in the presence of sampling. The initial low
performance showed in Fig. 4.3b is derived from the fact that we build the initial
classification library with the unsampled UPC-II trace. As a consequence, the
system needs to perform an initial retraining to build a representative model of
the current sampled scenario. As aforementioned, this also shows the importance of
the spatial obsolescence and justifies the importance of the Autonomic Retraining
System. Surprisingly, after the initial retraining, the system is able to sustain the
same accuracy as the unsampled scenario. The greater decrease of information
produced by packet sampling [46] is only reflected in a slight increment in the
number of retrainings given that, as described in Sec. 4.2.1, our techniques has
been adapted to deal with it.
Finally, in order to completely understand the influence of the Autonomic Retraining System, we have performed two additional experiments that confirm its
necessity. The first experiment creates a model with data from one network to classify traffic from another network (i.e., use the traceUPC-II to classify CESCA).
The second experiment creates the model with the traffic of the own network but
does not retrain it (i.e., use the CESCA trace to train and classify). Giving both
trainings can be performed offline, we used 3M of flows for both experiments,
instead of the 500K of our solution, trying to build the models as accurate and
representative as possible. Even so, Figs. 4.4 show that our solution outperforms
these experiments. Fig. 4.4a, that presents the results when no sampling is applied, shows two main outcomes. First, the origin from the data used in the
training impacts on the accuracy of the classification (i.e., spatial obsolescence).
Even both traces carry traffic from a similar scenario (i.e., educational/research
network) there is a substantial difference of accuracy as can be seen in the left
part of the figure. The second outcome arises in the right part of the figure where
both experiments obtain similar accuracy and this accuracy is gradually decreasing as long as times goes by (i.e., temporal obsolescence). On the other hand, our
solution keeps stable during the whole evaluation. Figure 4.4b, that presents the
results when 1/1 000 sampling rate is applied, emphasizes the outcomes previously
mentioned. Here, the application of sampling totally deprecates the classification
4.3. EVALUATION
69
100%
Accuracy
98%
96%
94%
Avg. accuracy = 96.76 % -- 5 retrainings -- 94% threshold
Avg. accuracy = 97.5 % -- 15 retrainings -- 96% threshold
Avg. accuracy = 98.26 % -- 108 retrainings -- 98% threshold
Fri, 04 Feb 2011
Tue, 08 Feb 2011
Fri, 11 Feb 2011
Mon, 14 Feb 2011
Thu, 17 Feb 2011
Time
(a) Retraining without sampling
100%
Accuracy
98%
96%
94%
Avg. accuracy = 96.65 % -- 5 retrainings -- 94% threshold
Avg. accuracy = 97.34 % -- 17 retrainings -- 96% threshold
Avg. accuracy = 98.22 % -- 116 retrainings -- 98% threshold
Fri, 04 Feb 2011
Tue, 08 Feb 2011
Fri, 11 Feb 2011
Mon, 14 Feb 2011
Thu, 17 Feb 2011
Time
(b) Retraining with 1/1 000 sampling rate
Figure 4.3: Impact of the Autonomic Retraining System on the Application Identifier with the selected configuration (i.e., naive training policy with 500K)
model created with the unsampled traffic of UPC-II producing a very inaccurate
classification. In both scenarios, the experiments that use CESCA for training and
classification start with a very high accuracy given that they are classifying the
same flows used for the training. Because of that, after the first 3M of flows the accuracy decreases even below than 86%. However, our solution with its continuous
retraining is able to deal with both temporal and spatial obsolescence achieving a
70
CHAPTER 4. AUTONOMIC TRAFFIC CLASSIFICATION SYSTEM
stable accuracy beyond 98%.
100%
98%
96%
Accuracy
94%
92%
90%
88%
Avg. accuracy = 92.73 % -- trained with UPC-II
Avg. accuracy = 94.3 % -- trained with first 3M CESCA flows
Avg. accuracy = 98.24 % -- 108 retrainings -- 98% threshold, naive training policy with 500K
Fri, 04 Feb 2011
Tue, 08 Feb 2011
Fri, 11 Feb 2011
Mon, 14 Feb 2011
Thu, 17 Feb 2011
Time
(a) Comparative without sampling
100%
98%
94%
90%
Accuracy
86%
82%
78%
74%
70%
66%
Avg. accuracy = 68.38 % -- trained with UPC-II
Avg. accuracy = 94.25 % -- trained with first 3M CESCA flows
Avg. accuracy = 98.22 % -- 116 retrainings -- 98% threshold, naive training policy with 500K
Fri, 04 Feb 2011
Tue, 08 Feb 2011
Fri, 11 Feb 2011
Mon, 14 Feb 2011
Thu, 17 Feb 2011
Time
(b) Comparative with 1/1 000 sampling rate
Figure 4.4: Comparative of the Autonomic Retraining System with other retraining
solutions
4.3.4
Retraining Evaluation by Institution
As described in Sec. 4.3.1, the CESCA trace was collected in the 10-Gigabit access
link of the Anella Cientı́fica, which connects the Catalan Research and Education
4.3. EVALUATION
71
100%
Accuracy
98%
96%
94%
92%
Avg. accuracy = 98.06 % -- 35 retrainings -- 98% threshold
Avg. accuracy = 98.41 % -- 108 retrainings -- 98% threshold
Fri, 04 Feb 2011
Tue, 08 Feb 2011
Fri, 11 Feb 2011
Mon, 14 Feb 2011
Thu, 17 Feb 2011
Time
(a) Institution A
100%
Accuracy
98%
96%
94%
92%
Avg. accuracy = 98.05 % -- 13 retrainings -- 98% threshold
Avg. accuracy = 97.91 % -- 109 retrainings -- 98% threshold
Fri, 04 Feb 2011
Tue, 08 Feb 2011
Fri, 11 Feb 2011
Mon, 14 Feb 2011
Thu, 17 Feb 2011
Time
(b) Institution B
100%
Accuracy
98%
96%
94%
92%
Avg. accuracy = 98.26 % -- 9 retrainings -- 98% threshold
Avg. accuracy = 98.17 % -- 108 retrainings -- 98% threshold
Fri, 04 Feb 2011
Tue, 08 Feb 2011
Fri, 11 Feb 2011
Mon, 14 Feb 2011
Thu, 17 Feb 2011
Time
(c) Institution C
Figure 4.5: Comparative of the Autonomic Retraining System by institution
Network with the Spanish Research and Education Network. This link provides
connection to Internet to more than 90 institutions. So far, the evaluation has been
performed using the complete traffic of the link. This section presents the results
72
CHAPTER 4. AUTONOMIC TRAFFIC CLASSIFICATION SYSTEM
of the performance of the Autonomic Retraining System with the disaggregated
traffic by institution.
Similarly to the previous evaluation, we have used 500K flows as training size
(i.e., X = 500K), the naive retraining policy (Sec. 4.3.2) and the highest accuracy threshold (i.e., 98%). Two different approaches are used in order to study
the performance by institution. First, the Autonomic Retraining System uses its
normal operation (i.e., using all the traffic and performing the retrainings based
on the total accuracy). However, only the accuracy related to the specific institution is presented. Second, the operation of the Autonomic Retraining System
is changed and, instead of using all the traffic, it only uses the traffic related to
the specific institution. Also, the decision of retraining is carried out based on
the particular accuracy of the institution and not with the total one. Figure 4.5
presents the results of this evaluation for three different institutions. The results
show the reliability of the Autonomic Retraining System for achieving high accuracies with different institutions and scenarios. Although the accuracy is very
similar between the three institutions, three different behaviors can be observed.
Institution A plotted in Fig. 4.5a has a very volatile accuracy. Even when the
model is trained with its own traffic the accuracy is sharply changing, although
almost always keeping an accuracy higher than 92%. On the other side, Institution C plotted in Fig. 4.5c has a more stable accuracy. This is translated into a
smaller number of trainings compared with Institution A. Finally, the behavior of
Institution B would be among the other two. These three behaviors plotted in
Figure 4.5 are the result of different grades of heterogeneity (i.e., Institution A)
and homogeneity (i.e., Institution C ) in the the traffic of the institutions.
Another interesting output from this evaluation is the impact of the origin of the
training data on the accuracy. Figure 4.5a shows that Institution A achieves higher
accuracy performing the retrainings with data from the complete link. However,
Fig. 4.5c shows that Institution C achieves higher accuracy when the classification
model is created with its own data. It is important to note that the amount of
retrainings is not comparable between the two approaches. Although the configuration is the same between them, the amount of data available for each approach
is different. These two different results regarding the two approaches could be
also related to the grade of heterogeneity of the traffic. Training the classification
model with traffic from others institutions can help to classify unexpected traffic
(e.g., new applications) in networks with heterogeneous traffic.
All these results confirm that the combination of the three techniques and the
ability to automatically update the classification model outperform the solutions
proposed in the literature for Sampled NetFlow traffic classification [15, 46]. The
4.4. RELATED WORK
73
proposed system has been deployed in production in the Catalan Research and
Education network and it is currently being used by network managers of more
than 90 institutions connected to this network.
4.4
Related Work
Although some works have also proposed the combination of different classification
techniques [88–90], previous solutions do not support sampling, require packetlevel data and cannot automatically adapt the classification model to the changing
conditions of the network traffic and applications. To the best of our knowledge,
only one previous paper has addressed the problem of automatic retraining [64].
However, it only presents a superficial evaluation with the unique goal of showing
the feasibility of retraining with a classifier based on K-dimensional trees. In
this work, we implement and test a complete automatic retraining system able to
update multiple classification techniques without relying on human intervention.
4.5
Chapter Summary
In this chapter, we presented a realistic solution for traffic classification for network
operation and management. Our classification system combines the advantages
of three different techniques (i.e., IP-based, Service-based and ML-based) along
with an autonomic retraining system that allows it to sustain a high classification
accuracy during long periods of time. The retraining system combines multiple
DPI techniques and only requires a small sample of the whole traffic to keep the
classification system updated.
Our experimental results using a long traffic trace from a large operational
network shows that our system can sustain a high accuracy (>96%) and completeness during weeks even with Sampled NetFlow data. We also evaluated different
training policies and studied their impact on the traffic classification technique.
From these results we can draw several conclusions:
• The classification models obtained suffer from temporal and spatial obsolescence. Our results in Sec. 4.3.3 confirm this problem which was already
pointed out by Li et al. in [47]. Our system addressed this problem by implementing the Autonomic Retraining System that is able to automatically
update the classification models without human supervision.
74
CHAPTER 4. AUTONOMIC TRAFFIC CLASSIFICATION SYSTEM
• The life of the classification models is not fixed. As indicated by the results
in Sec. 4.3.4, we show that the frequency of retrainings partially depends on
the grade of heterogeneity and volatility of the traffic in the network.
• Although several classification techniques have been proposed by the research
community, there is no one suitable for all the types of traffic and scenarios. We truly believe that the combination of different techniques is the
best approach for properly classifying all the different types of traffic. Our
approach, based on three different techniques, is able to achieve very high
accuracy and completeness, something that would not be possible if they
were not combined.
In summary, we presented a traffic classification system with several features
that are particularly appealing for network management: (i) high classification
accuracy and completeness, (ii) support for NetFlow data, (iii) automatic model
retraining, and (iv) resilience to sampling. These features altogether result in a
significant reduction in the cost of deployment, operation and maintenance compared to previous methods based on packet traces and manually-made classification
models. The proposed system has been deployed in production in the Catalan Research and Education network and it is currently being used by network managers
of more than 90 institutions connected to this network.
Chapter 5
Streaming-based Traffic
Classification System
5.1
Introduction
The main problem of previous proposals, like the one presented in Chapter 3, is
that most of these techniques are based on a static view of the network traffic (i.e.
they build a model or a set of patterns from a static, invariable dataset). However,
very few works have addressed the practical limitations that arise when facing a
more realistic scenario with an infinite, continuously evolving stream of network
traffic flows.
On the contrary, this chapter proposes a flow-based network traffic classification solution that can automatically adapt to the continuous changes in the
network traffic. We introduce for the first time the use of Hoeffding Adaptive
Tree (HAT) for traffic classification. In contrast to previous solutions that rely
on static datasets, this technique addresses the classification problem from a more
realistic point of view, by considering the network traffic as an evolving, infinite
data stream. This technique has very appealing features for network traffic classification, including the following:
• It processes a flow at a time and inspects it only once (in a single pass), so
it is not necessary to store any traffic data.
• It uses a limited amount of memory, independent of the length of the data
stream, which is considered infinite.
75
76
CHAPTER 5. STREAMING-BASED TRAFFIC CLASSIFICATION
• It works in a limited and small amount of time, so it can be used for online
classification.
• It is ready to predict at any time, so the model is continuously updated and
ready to classify.
Our solution also has some interesting features that simplify its deployment in
operational networks compared to other alternatives based on DPI or ML techniques. The main problem with DPI-based techniques is that they rely on very
powerful and expensive hardware to deal with nowadays traffic loads, which must
be installed in every single link to obtain a full coverage of a network [10, 43, 44].
Similarly, traditional ML-based techniques for traffic classification require access
to individual packets, which involves the use of optical splitters or the configuration of span ports in switches. In contrast, our solution works at the flow level and
is compatible with NetFlow v5, a widely extended protocol developed by Cisco
to export IP flow information from network devices [77], which has already been
deployed in most routers and switches. Although our solution uses NetFlow v5 as
input, it can easily work with other similar exporting protocols (e.g., IPFIX [55]).
In order to present sound conclusions about the quality, simplicity, and accuracy of our proposal, we evaluate our traffic classification solution with the entire
MAWI dataset [59] described in Section 2.3.3, a unique publicly available dataset
that covers a period of 13 years. To the best of our knowledge, this is the first work
that deals with such amount of real traffic data for traffic classification. Our results show that our solution for traffic classification is able to automatically adapt
to the changes in the traffic over the years, while sustaining very high accuracies.
We show that our technique is not only more accurate than other state-of-the-art
techniques when dealing with evolving traffic, but it is also less complex and easy
to maintain and deploy in operational networks.
The rest of this chapter is organized as follows. The proposed classification
technique based on Hoeffding Adaptive Trees is described in Section 5.2. The
methodology used for the evaluation of our technique is presented in Section 5.3.
Section 5.4 analyzes the impact of different configuration parameters of HAT when
used for network traffic classification. Section 5.5 evaluates our solution based on
HAT with the MAWI dataset and compares it with the decision tree C4.5 [71], a
widely used supervised learning technique. The related work is briefly presented
in Section 5.6. Finally, Section 5.7 concludes the work.
5.2. CLASSIFICATION OF EVOLVING NETWORK DATA STREAMS
5.2
77
Classification of Evolving Network Data
Streams
We propose a flow based traffic classification technique for evolving data streams
based on Hoeffding Adaptive Trees. This technique has very interesting features
for network traffic classification, and addresses the classification problem from a
more realistic point of view, because it considers the network traffic as a stream
of data instead of as a static dataset. This way, we better represent the actual
streaming-nature of the network traffic and address some practical problems that
arise when these techniques are deployed in operational networks. We describe
our proposal to classify network traffic streams in this section. We first present
the original Hoeffding Tree (HT) technique oriented to data streams and then we
briefly describe the adaptation to deal with evolving data streams, called Hoeffding
Adaptive Tree (HAT). Finally, we present the traffic attributes selected to perform
the classification of the network traffic.
5.2.1
Hoeffding Tree
Hoeffding Tree (HT) is a decision tree-based technique oriented to data streams
originally introduced by Hulten et al. in [91]. As already mentioned, streamoriented techniques have many appealing features for network traffic classification:
(i) they process an example at a time and inspect it only once (i.e., they process the
input data in a single pass), (ii) they use a limited amount of memory independent
of the length of the data stream, which is considered infinite, (iii) they work in a
limited amount of time, and (iv) they are ready to predict at any time. However,
these features considerably complicate the induction of the classification model.
ML batch techniques (e.g., C4.5, Naive Bayes) are usually performed over static
datasets, and therefore, they have access to the whole training data to build the
model as many times as needed. On the contrary, models resulting from streamoriented techniques should be inducted incrementally from the data they process
just once and on-the-fly. Therefore, the technique cannot store any data related
to the training, which makes the decision-making a critical task.
A key operation in the induction of a decision tree is to decide when to split
a node. Batch techniques have access to all the data in order to perform this
operation and decide the most discriminating attribute in a node. HT uses the
Hoeffding bound [92] in order to incrementally induce the decision tree. Briefly,
this bound guarantees that the difference of discriminating power between the best
78
CHAPTER 5. STREAMING-BASED TRAFFIC CLASSIFICATION
attribute and the second best attribute in a node can be well estimated if enough
instances are processed. The more instances it processes the smaller is the error.
The method to compute this discriminating power, which depends on the split
criteria (e.g., Information Gain), as well as other HT parameters are later studied
in Sec. 5.4.
5.2.2
Hoeffding Adaptive Tree
Hoeffding Tree allows the induction of a classification model according to the
requirements of a data stream scenario. However, an important characteristic
of the Internet is that the stream of data continually changes over time (i.e.,
it evolves). Batch models should be regularly retrained in order to adapt the
classification model to the variations of the network traffic, which is a complex and
very costly task [61]. Hoeffding Adaptive Tree (HAT), proposed in [93], solves this
problem by implementing the Adaptive Sliding Window (ADWIN). This sliding
window technique is able to detect changes in the stream (i.e., concept drift) and
provide estimators of some important parameters of the input distribution using
data saved in a limited and fixed amount of memory, which is independent of the
total size of the data stream. The interested reader is referred to [94] for more
details on how ADWIN is implemented.
5.2.3
Inputs of our System
The implementation of our system can indistinctly receive two different types
of instances: labeled and unlabeled flows. Depending on the type of instance,
our solution will perform a classification (if the flow is not labeled) or a training
operation (if it is labeled). The classification process labels a new unknown flow
using the HAT model. The input of the classification process consists of a set of
16 flow features that can be directly obtained from NetFlow v5 data: source and
destination port, protocol, ToS, # packets, # bytes, TCP flags, average packet
size, flow time, flow rate, and flow inter-arrival time. The choice of features is based
on our previous work in [61]. The use of standard NetFlow v5 data considerably
decreases the cost of deployment and computation requirements of the solution,
given that the input is already provided directly by the routers.
The other type of instances our solution can receive are the retraining flows.
These flows will be labeled by an external tool, as will be described later. In order
to automatically update the model, our technique should receive training flows
with the same set of 16 features used by the classification process together with
5.3. METHODOLOGY
79
the label associated to them. Unlike batch techniques, the retraining process is
performed incrementally, which allows the model to be ready to classify at any
time. Therefore, our solution can indistinctly deal with a mix of instances and
operate with them according to their type (i.e., classification or retraining flows).
The best ratio between classification and retraining instances depends on the scenario to be monitored. However, as shown in [61], a very small ratio of retraining
instances (e.g., less than 1/400) is sufficient to keep a high accuracy along time.
This labeling process can be performed with several techniques, including DPI,
given that only a small sample of the traffic needs to be labeled, and therefore
it is computationally lightweight. For instance, a common example would be the
deployment of our solution in a network with several routers exporting NetFlow
v5 data. The labeling of the training flows could be done with NBAR2 [95], using
a small sample of the traffic from only one the routers. NBAR2 is a DPI-based
technique implemented in the last versions of the Cisco IOS. Otherwise, activating
NBAR2 in all the routers and with all the traffic is usually not possible, given the
high computational cost and impact it would have on their performance. Another
alternative is the use of the methodology presented in [61]. This consists of a small
sample of data with full payload, which is labeled using an external DPI tool. This
is the solution used in the evaluation presented in Sec. 5.5.
5.3
Methodology
This section describes the methodology used to evaluate the performance of our
proposal. First, the tool used for the evaluation is presented and then, the dataset
used as ground-truth for the evaluation is described.
5.3.1
MOA: Massive Analysis Online
Massive Online Analysis (MOA) [96] is a Java open source software for data stream
mining. Unlike its well-known predecessor WEKA [72], MOA is oriented to the
evaluation and implementation of ML techniques for data streams. It is specially
designed to compare the performance of stream oriented techniques in streaming scenarios. MOA implements the HAT technique with a set of configuration
parameters. In addition, it allows the use of batch techniques implemented in
WEKA, which simplifies the comparison of traditional batch ML techniques like
the decision tree C4.5.
MOA implements different benchmark settings to evaluate stream techniques.
80
CHAPTER 5. STREAMING-BASED TRAFFIC CLASSIFICATION
For our evaluation, we chose Evaluate Interleaved Chunks among the different
options available in MOA. Interleaved Chunks uses all the instances dividing the
stream in chunks (i.e., set of instances). Every chunk is used first for testing and
then for training.
We believe that this approach is the most representative because it uses the
complete dataset (i.e., stream) for both testing and training. Similar conclusions
are drawn with other evaluations methods. In our evaluation we first use the
default configuration of the Interleaved Chunks parameters to simplify its comparison. We then study the impact of the chunk size on its performance.
5.3.2
The MAWI Dataset
In order to obtain representative results for the evaluation of stream oriented techniques, we need datasets that are long enough to capture the evolution of Internet
traffic over time. We use the publicly available MAWI dataset [59] to perform
the evaluation because it has unique characteristics to study stream oriented techniques for network traffic classification. Although it is a static dataset, its long
duration (i.e., 13 years) and amount of data makes it the perfect candidate for
the evaluation of our technique. Furthermore, its duration allows us to study the
ability of HAT to automatically adapt to the evolution of the traffic. Section 2.3.3
describes in detail the MAWI dataset, the methodology used to obtain the groundtruth and its traffic mix. In our evaluation, we performed a sanitization process
and focused on the TCP and UDP traffic from the MAWI dataset. After the labeling and the sanitization process, the MAWI dataset consists of almost 4 billions
of unidirectional labeled flows. To the best of our knowledge, this is the first work
in the network traffic classification field that deals with this large amount of data,
which is necessary to extract sound conclusions from our evaluation.
5.4
Hoeffding Adaptive Tree Parametrization
In this section, we study the parametrization of the Hoeffding Adaptive Tree technique for network traffic classification. As described in Section 5.3, we use MOA
and the MAWI dataset to perform the evaluation. Since this is the first work to
use HAT for network traffic classification, the configuration of the different parameters of HAT and their impact on network traffic classification remain unknown.
Because of this, we next present a complete study of the impact of the different
parameters of HAT when applied to network traffic. We have studied a total of ten
5.4. HOEFFDING ADAPTIVE TREE PARAMETRIZATION
81
parameters implemented in MOA for HAT: numeric estimator, grace period, tie
threshold, split criteria, leaf prediction, stop memory management, binary splits,
remove poor attributes, no preprune, and split confidence. In this section, we chose
40 million of instances to perform the evaluation. We split them in four different
dates to ensure the representativeness of the results, more exactly we have selected
the first 10 million of instances from October 2001, January 2004, July 2008 and
March 2011. We perform a specific experiment for each date and then compute
the average of them to present the results. After the parametrization, Section 5.5
presents an evaluation with the complete MAWI dataset. We briefly describe each
parameter, however, we refer the interested reader to [93] for a detailed explanation.
5.4.1
Numeric Estimator
An important issue of ML techniques oriented to data streams is how they deal with
numeric attributes. Unlike most batch ML techniques (e.g., C4.5, Naive Bayes),
the techniques for data streams can only pass one time over the data. Because
of that, the discretization of the features (i.e., numeric attributes are transformed
into discrete attributes) is a more difficult task. MOA implements 4 different
numeric estimators for classification using HAT: Exhaustive Binary Tree, Very Fast
Machine Learning (VFML), Gauss Approximation (i.e., default one), and Quantile
Summaries (i.e., Greenwald-Khanna). Figure 5.1 (top) presents the performance
results of this criteria. We tested different values for each numeric estimator,
however, we studied more values of the VFML numeric estimator given its better
results. These values correspond to the number of bins used for discretization of the
numeric attributes. Gauss Approximation as much as Greenwald-Khanna obtain
very poor results. The best numeric estimators in our scenario are VFML and
the Exhaustive Binary Tree. More specifically, VFML 1 000 and the Exhaustive
Binary Tree are the most accurate.
Apart from the accuracy, another important feature to take into account is the
overhead every option implies. Note that this technique should work online and
deal with a huge amount of data in a limited amount of time. Because of this, it
is important to keep the solution as lightweight as possible while keeping a high
accuracy. Figure 5.1 (bottom) presents the model cost of each numeric estimator
in our evaluation. Greenwald-Khanna, Gauss Approximation, and VFML 10 and
100 are hidden behind VFML 1 000. The huge difference of load between the three
most accurate techniques makes the VFML 1 000 the best numeric estimator for
our scenario.
82
CHAPTER 5. STREAMING-BASED TRAFFIC CLASSIFICATION
100
Evaluate Interleaved Chunks HAT
% Accuracy
80
60
40
20
Cost (Gb per hour)
0
0.0
4.5
4.0
3.5
3.0
2.5
2.0
1.5
1.0
0.5
0.00.0
VFML 10
VFML 100
VFML 1000
VFML 10000
BT
GREEN (10,100)
GAUSS (10,100)
0.2
0.4
0.6
# Flows
0.8
1.0
1e7
Evaluate Interleaved Chunks HAT
VFML 10
VFML 100
VFML 1000
VFML 10000
BT
GREEN (10,100)
GAUSS (10,100)
0.2
0.4
0.6
# Flows
0.8
1.0
1e7
Figure 5.1: Impact of the Numeric Estimator parameter on the HAT technique
5.4.2
Grace Period
The next parameter studied is the grace period. This parameter configures how
often (i.e., how many instances between computations) the values in the leafs of
HAT are computed. This computation is performed in order to decide if a further
split is necessary. This computation is considerably costly and the impact of each
instance in the result of this computation is small. Therefore, it is reasonable
to perform this computation periodically instead of repeating it for each instance.
High values would reduce the cost of the technique but slow down the growth of the
tree, thus decreasing its accuracy in theory. Figure 5.2 (top) presents the impact
5.4. HOEFFDING ADAPTIVE TREE PARAMETRIZATION
83
of different grace values on the accuracy of the technique. At first glance, there
are no huge differences between the different values. As expected, the lowest value
is initially getting the best results since it is extracting the knowledge by quickly
splitting the leaves. However, we are dealing with a data stream and making a
decision with few instances can sometimes produce inaccuracies in the future. In
Fig. 5.2 (top), the most accurate grace periods are 1 000 and 200 (i.e., default
one). Both values are able to keep a stable high accuracy and avoid down peaks
presented for the other values.
However, the importance of this parameter is its ability to decrease the overhead of the technique without decreasing significantly its accuracy. Figure 5.2
(bottom) presents how the different values of the grace period affects to the cost
of the technique. We decided to use 1 000 as grace period giving it is the best
trade-off between accuracy and load.
5.4.3
Tie Threshold
A well-known parameter from decision tree techniques is the tie threshold. Sometimes two or more attributes in a leaf cannot be separated because they have
identical values. If those attributes are the best option for splitting the node, the
decision would be postponed until they differ and this can decrease the accuracy.
Figure 5.3 (top) presents the accuracy obtained with different values of the tie
threshold parameter. The most accurate value is 1, closely followed by 0.5 and
0.25.
In order to decide between the most accurate tie thresholds, we rely on the
cost of the model they produce. Figure 5.3 (bottom) shows that 0.25 and 1 are the
best options depending on the evaluation approach among the three more accurate
values. We decided to use 1 as tie threshold because it is the most accurate.
5.4.4
Split Criteria
As mentioned before, the grace period indicates when to compute the necessary
values to decide if a node should be split. This computation refers to the split
criteria. This parameter decides when an attribute is enough discriminative to
split a node. There are two approaches implemented in MOA: Information Gain
and Gini. Figure 5.4 (top) presents the accuracy obtained with the Gini split
criteria and different values of the Information Gain. These values correspond
to the minimum fraction of weight required to down at least two branches. The
performance of the Gini option is considerably poor in our scenario. Regarding
84
CHAPTER 5. STREAMING-BASED TRAFFIC CLASSIFICATION
100
Evaluate Interleaved Chunks HAT
% Accuracy
80
60
40
20
Cost (Gb per hour)
0
0.0
0.16
0.14
0.12
0.10
0.08
0.06
0.04
0.02
0.000.0
Grace 5000
Grace 2000
Grace 1000
Grace 200
Grace 50
0.2
0.4
0.6
# Flows
0.8
1.0
1e7
Evaluate Interleaved Chunks HAT
Grace 5000
Grace 2000
Grace 1000
Grace 200
Grace 50
0.2
0.4
0.6
# Flows
0.8
1.0
1e7
Figure 5.2: Impact of the Grace Period parameter on the HAT technique
the different values of the Information Gain, the values 0.001, 0.01, and 0.1 achieve
the highest accuracies.
Figure 5.4 (bottom) shows how the cost of technique is impacted by the different
split criteria. We decided to use the Information Gain value 0.001 because it is
the lightest among the most accurate.
5.4.5
Leaf Prediction
An important feature of HAT is that, since the model is continuously being updated, it is always ready to classify. The next parameter is related to this classi-
5.4. HOEFFDING ADAPTIVE TREE PARAMETRIZATION
100
85
Evaluate Interleaved Chunks HAT
% Accuracy
80
60
40
20
Cost (Gb per hour)
0
0.0
0.040
0.035
0.030
0.025
0.020
0.015
0.010
0.005
0.0000.0
tie 1
tie 0.5
tie 0.25
tie 0.1
tie 0.05
tie 0.001
0.2
0.4
0.6
# Flows
0.8
1.0
1e7
Evaluate Interleaved Chunks HAT
tie 1
tie 0.5
tie 0.25
tie 0.1
tie 0.05
tie 0.001
0.2
0.4
0.6
# Flows
0.8
1.0
1e7
Figure 5.3: Impact of the Tie Threshold parameter on the HAT technique
fication and describe how HAT performs the classification decision at leaf nodes.
MOA implements three different approaches: Majority Class, Naive Bayes and
Naive Bayes Adaptive. The Majority Class approach consists of assigning the
most frequent label in that leaf. Apart from the most frequent label in a leaf, we
have much information related to the instance (i.e., attributes). The Naive Bayes
approach tries to use this extra information to make a more accurate prediction.
This approach computes the probability an instance belongs to the different possible labels from a leaf based on its attributes. The most probable label is the
one assigned. However, this technique can reduce the accuracy depending on the
scenario. The Naive Bayes Adaptive approach tries to take advantage of both
86
CHAPTER 5. STREAMING-BASED TRAFFIC CLASSIFICATION
100
Evaluate Interleaved Chunks HAT
% Accuracy
80
60
40
20
Cost (Gb per hour)
0
0.0
0.040
0.035
0.030
0.025
0.020
0.015
0.010
0.005
0.0000.0
InfoGain 0.001
InfoGain 0.01
InfoGain 0.1
InfoGain 0.25
InfoGain 0.5
Gini
0.2
0.4
0.6
# Flows
0.8
1.0
1e7
Evaluate Interleaved Chunks HAT
InfoGain 0.001
InfoGain 0.01
InfoGain 0.1
InfoGain 0.25
InfoGain 0.5
Gini
0.2
0.4
0.6
# Flows
0.8
1.0
1e7
Figure 5.4: Impact of the Split Criteria parameter on the HAT technique
approaches by combining them. It computes the error rate of the Majority Class
and Naive Bayes in every leaf, and use for future predictions the approach that
has been more accurate so far. Figure 5.5 (top) presents the accuracy obtained
with the different approaches. Unexpectedly, the Naive Bayes approach obtains
very poor results. As described in [97], the experimental implementation in MOA
does not change the memory management strategy when Naive Bayes is enabled
and this can impact on its performance. On the other hand, the Majority Class
and the Naive Bayes Adaptive approaches obtain similar high accuracies.
Figure 5.5 (bottom) shows how the different approaches impact the solution
in terms of model cost. Taking into account these results, we decided to use
5.4. HOEFFDING ADAPTIVE TREE PARAMETRIZATION
87
Majority Class as the leaf prediction technique. Apart from having a lower cost,
while achieving similar high accuracy, the Majority Class approach is not affected
by other parameters. Approaches based on Naive Bayes can decrease its accuracy
if parameters like removing poor attributes or stopping memory management are
activated.
100
Evaluate Interleaved Chunks HAT
% Accuracy
80
60
40
20
0
0.0
0.030
Cost (Gb per hour)
0.025
Majority Class
Naive Bayes
Naive Bayes Adaptive
0.2
0.4
0.6
# Flows
0.8
1.0
1e7
Evaluate Interleaved Chunks HAT
Majority Class
Naive Bayes
Naive Bayes Adaptive
0.020
0.015
0.010
0.005
0.0000.0
0.2
0.4
0.6
# Flows
0.8
1.0
1e7
Figure 5.5: Impact of the Leaf Prediction parameter on the HAT technique
5.4.6
Other Parameters
So far, the parameters studied have substantially impacted the accuracy or cost
of HAT. However, we have also evaluated some parameters with marginal impact.
88
CHAPTER 5. STREAMING-BASED TRAFFIC CLASSIFICATION
This is the case of the Stop Memory Management parameter. When this parameter is activated, HAT stops growing as soon as the memory limit is reached.
However, it seems that the default value of the memory limit in MOA is never
reached or this parameter is not implemented for the HAT technique. The Binary
Split parameter, describing if the splits of a node have to be binaries or not, has
also a marginal impact. We truly believe that this result is directly related to
our scenario characteristics. All our attributes are numerically and hence all the
splits performed are almost always binary splits. The last parameter studied with
marginal impact is the Remove Poor Attributes parameter. This feature removes
attributes in the leafs whose initial values indicate their uselessness for the splitting decision. In our scenario, these parameters have not impacted the accuracy
of HAT. However, a marginal improvement has been observed in terms of cost.
Thus, we also activated them in the final configuration.
We have also studied the parameters No PrePrune and Split Confidence and
no differences have been observed. As a result, none of them are activated in our
final configuration.
Finally, similarly to other ML-based techniques, HAT can be used in ensembles
techniques. MOA implements several ensembles methods (e.g., bagging, boosting)
that basically combine several models to improve the final accuracy. However,
this improvement comes with a higher computational cost. Given that we already
achieve a very high accuracy with the current configuration, we dismissed the use
of ensembles techniques in our scenario.
Table 5.1 presents the final configuration of the parameters obtained in this
section. We use this configuration for the evaluation of the HAT technique for
network traffic classification.
5.5
Hoeffding Adaptive Tree Evaluation
Once the best configuration is selected we compare the HAT technique with a wellknown technique from the literature. The goal of this comparison is to show that
our solution can be as accurate as batch-oriented techniques, but with the appealing features of those oriented to streams. As mentioned in Sec. 5.1, batch techniques are usually built from a static dataset and do not address the ever-changing
nature of the Internet traffic [47] or rely on complex custom-made solutions [61].
However, our solution can automatically adapt to changing traffic conditions without storing any data and being always ready to classify. For this comparison, we
chose the J48 technique as a representative example of batch-oriented techniques,
5.5. HOEFFDING ADAPTIVE TREE EVALUATION
89
Table 5.1: HAT parametrization for network traffic classification
Parameter
Numeric Estimator
Grace Period
Tie Threshold
Split Criteria
Leaf Prediction
Stop Memory Management
Binary Splits
Remove Poor Attributes
Value
VFML with 1 000 bins
1 000 instances (i.e., flows)
1
Information Gain with 0.001
as minimum fraction of weight
Majority Class
Activated
Activated
Activated
which is an open source version of the C4.5 decision tree implemented in WEKA.
We selected this technique because it has been widely used for network traffic
classification [37,46,47,61], achieving very good results when compared with other
techniques [32, 36].
5.5.1
Single Training Evaluation
Usually, ML-based network traffic classification solutions presented in the literature are evaluated from a static point of view using limited datasets. The first
evaluation performed pretends to show the temporal obsolescence of the models
produced with static datasets [47, 61]. To achieve this goal, we performed an
evaluation applying just an initial training with 3 million of flows in 2001 for the
complete classification of the 13 years of traffic of the MAWI dataset. Figure 5.6
presents how the accuracy of both techniques is substantially degraded in this
evaluation showing that the models should be regularly updated to adapt to the
changes in the traffic. The deep drops in the accuracy are related to new applications that are not present in the initial training dataset. The increment of accuracy
during the last years of the evaluation is due to the change of the traffic mix in
the MAWI dataset. As showed in Table 2.5, there is an increment of traditional
applications (i.e., DNS, HTTP and NTP) and a decrease of novel applications (i.e.,
BitTorrent and Skype) during those years. Given that this evaluation is performed
from a static point of view, HAT is not able to make use of its interesting features
for streams.
90
CHAPTER 5. STREAMING-BASED TRAFFIC CLASSIFICATION
100
Single Training Evaluation
% Accuracy
80
60
40
20
HAT
J48
0 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013
Time
Figure 5.6: HAT comparison with single training configuration
5.5.2
Interleaved Chunk Evaluation
The second experiment consists of an Interleaved Chunk evaluation with the default evaluation method of MOA. That is, a stream-based evaluation where the
4 000 million of flows from the 13 years of the MAWI dataset are segmented in
chunks of 1 000 instances that are first used to classify and later to train. Figure 5.7
presents the results regarding this evaluation. Our solution achieves considerably
better results than the J48 batch technique. This can be easily explained by
the fact that this evaluation methodology is oriented to evaluate incrementally
inducted techniques. The J48 batch technique creates a new decision tree from
scratch with every chunk of 1 000 instances forgetting all the previous knowledge
extracted. In contrast, our solution updates the classification model with the new
information but considering also all the information extracted so far, which results
in a more robust classification model.
5.5.3
Chunk Size Evaluation
As shown in the previous experiment, J48 is significantly less accurate than HAT
with the default stream-based evaluation. However, that difference seems mainly
5.5. HOEFFDING ADAPTIVE TREE EVALUATION
100
91
Interleaved Chunks Evaluation
% Accuracy
80
60
40
20
HAT
J48
0 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013
# Flows
Figure 5.7: Interleaved Chunk evaluation with default configuration
because of the small chunk size that produces very poor J48 trees. In order to
address this problem, we next study the impact of the chunk size on both techniques. We evaluate six different chunk sizes (i.e., 1, 100, 1 000, 10 000, 100 000,
1 000 000 flows) in the Interleaved Chunk evaluation. Given the large number of
executions involved, we decided to use a sample of more than 4 million of flows
of the MAWI dataset in this experiment. Figure 5.8 shows the accuracy of both
techniques for each chunk size. Given that HAT builds its tree incrementally, it is
barely affected by the chunk size, achieving always a very high accuracy. Unlike
HAT, J48 is substantially impacted by the chunk size. As expected, the small
values of the chunk size (i.e., 1, 100, 1 000) produce inaccurate J48 trees. Only the
highest chunk sizes (i.e., 100 000 and 1 000 000) are able to achieve similar accuracies to the HAT technique. Moreover, large chunk sizes imply the storage of large
amounts of traffic as we will discuss next.
As important as the accuracy is the cost of the techniques. The J48 decision
tree, as a batch technique, needs to store first the data of each chunk to continuously build the model from scratch, which results in huge memory requirements.
Figure 5.9 presents the cost (i.e., bytes per second) by flow in log scale directly
obtained from MOA. For clarity, only the extremes values (i.e., 1, 1 000 000) and
92
CHAPTER 5. STREAMING-BASED TRAFFIC CLASSIFICATION
100
Interleaved Chunks by Chunk Size
% Accuracy
80
60
40
20
0
100
HAT
J48
101
102
103 104
Chunk Size
105
106
Figure 5.8: HAT accuracy comparison by chunk size
the default value (1 000) are plot. The rest of values follow a similar behavior as
the 1 000 chunk size. Initially, all the sizes have a high cost per flow, especially the
smallest and the highest chunk sizes (i.e., 1 and 1 000 000). The cost quickly decreases after the initial peak. However, it decreases differently for both techniques.
After the initial peak, the cost of J48 remains more or less constant along time.
The cost for J48 among the different chunk sizes is similar but the highest chunk
size (i.e., 1 000 000), being more than five times higher. In contrast, the cost of
HAT rapidly decreases to very low values. Even with the highest chunk size, it is
able to decrease the cost similarly to the lowest values of the J48 technique. The
constant cost of J48 is related to the cost of the training of each model for each
chunk. Unlike J48, the model of HAT is incrementally built. Once it is consistent
(i.e., around 2 million in our evaluation) only small modifications are applied in
the model for every chunk.
To better show the differences in the cost of both techniques, Figure 5.10
presents the accumulated cost of both techniques by chunk size. The growth of
the cost by the HAT technique is almost plain after 2 million of flows. On the other
hand, J48 has a continuous growth along time. It is important to note that this
evaluation is done with a static dataset of 4 million. However, the difference of cost
5.5. HOEFFDING ADAPTIVE TREE EVALUATION
Cost (bytes per second)
10
1
93
Interleaved Chunks by Chunk Size
Cost by Flow
100
10-1
HAT_1000000
HAT_1000
HAT_1
J48_1000000
J48_1000
J48_1
10-2
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0
1e7
# Flows
Figure 5.9: HAT cost comparison by chunk size
between both techniques would considerably increase in an infinite stream-based
scenario (e.g., network traffic classification).
In summary, in a stream-based scenario, the HAT technique is usually more
accurate than J48. Only when high chunk sizes are used, J48 is able to be as
accurate as the HAT technique. Furthermore, HAT consumes less resources than
the J48 decision tree, especially when those high chunk sizes (i.e., 100 000 and
1 000 000) are used to increase the accuracy of J48.
5.5.4
Periodic Training Evaluation
In order to compare our results with other retraining proposals from the literature,
we modified the original idea of the Interleaved Chunk evaluation by following
the configuration proposed in [61]. The new evaluation consists of the use of
chunks of 500 000 instances for training and 500 000 000 for testing, using the
last seen chunk to train the next group. This evaluation represents the scenario
presented in Section 5.2.3, where a sample of the traffic is labeled by a DPIbased technique to retrain the model, while it is used to classify all the traffic.
Therefore, with the exception of the first chunk, the complete MAWI dataset is
94
CHAPTER 5. STREAMING-BASED TRAFFIC CLASSIFICATION
1.2
Cost (Gb per hour)
1.0
0.8
Interleaved Chunks by Chunk Size
Accumulated Cost
HAT_1000000
HAT_1000
HAT_1
J48_1000000
J48_1000
J48_1
0.6
0.4
0.2
0.00.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0
1e7
# Flows
Figure 5.10: HAT accumulated cost comparison by chunk size
classified. We selected 500 000 as chunk size derived from the results obtained
in [61]. However, in [61], the retrained decision is based on a threshold accuracy
while, in our evaluation, due to software constraints, it is based on the amount of
instances processed (i.e., 500 000 000). Although the evaluation has been changed,
the operation to compute the accuracy is maintained to make the comparison
possible. Figure 5.11 presents the results of this evaluation. The accuracy of the
J48 technique has been improved significantly. However, the stable accuracy seen
in the previous evaluation has changed to a more volatile one. This is because
the initial configuration is continuously retrained and quickly adapting itself to
the changes in the traffic. The results suggest that in this particular dataset, the
retraining should be performed more often in order to adapt faster to the changes
in the traffic with the related cost it would produce. Note, however, that the choice
of the chunk size for HAT is quite irrelevant, as shown in Section 5.5.3
5.5.5
External Evaluation
So far, we presented the parametrization and evaluation of the HAT technique
with the MAWI traffic. The results show that the HAT technique is, at least,
5.5. HOEFFDING ADAPTIVE TREE EVALUATION
100
95
Periodic Training Evaluation
% Accuracy
80
60
40
20
HAT
J48 [8]
0 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013
Time
Figure 5.11: Interleaved Chunk comparison with [61] configuration
as accurate as a state-of-the-art technique, such as C4.5 (i.e., J48 in MOA) but
with considerably less cost. In order to show these results are not only related
to the MAWI dataset, we next evaluate the performance of the HAT technique
with a different dataset. We used the CESCA dataset used in [61] to compare
the performance of HAT and J48 and make easier the comparison between both
works. The CESCA dataset, described in detail in Section 2.3.2, is a fourteendays packet trace collected on February 2011 in the 10-Gigabit access link of the
Anella Cientı́fica, which connects the Catalan Research and Education Network
with the Spanish Research and Education Network. A 1/400 flow sampling rate
was applied accounting for a total of 65 million of labeled flows. We use the
default configuration of the Interleaved Chunk evaluation and the parametrization
obtained in Section 5.4 for the HAT configuration. Although possible tuning could
be applied to this specific scenario, we show that the configuration obtained in
Section 5.4 seems suitable for other scenarios. Figure 5.12 shows that, similar to
Fig. 5.7, the HAT technique is more accurate than J48 in a stream-based scenario.
The smaller differences in terms of accuracy with the CESCA dataset can be
related to a less heterogeneous traffic mix and a shorter dataset (i.e., 14 days vs
13 years).
96
CHAPTER 5. STREAMING-BASED TRAFFIC CLASSIFICATION
Interleaved Chunks Evaluation
100
% Accuracy
80
60
40
20
00
HAT
J48
1
2
4
3
# Flows
5
6
1e7
Figure 5.12: Interleaved Chunk evaluation with CESCA dataset
5.6
Related Work
Machine Learning techniques for evolving data streams have been widely used in
many fields during the last years [98,99]. However, their application in the field of
network traffic classification has been minimal despite of their appealing features.
To the best of our knowledge just two authors have applied similar techniques in the
network traffic classification field. Tian et al. in [56,57] presented an evaluation of a
tailor-made technique oriented to data streams. They also compare it with different
ML batch techniques from the literature (i.e., C4.5, BayesNet, Naive Bayes and
Multilayer Perceptron). The results obtained are aligned with our results, however,
the dataset used was very limited for the evaluation of a stream data technique (i.e.,
2 000 instances per application). Also, Raahemi et al. introduced in [58] the use
of Concept-adapting Very Fast Decision Tree [91] for network traffic classification.
This technique, closely related to HAT, achieves high accuracy. However, the study
focus on the differentiation of P2P and non-P2P traffic. The dataset was labeled
using a port-based technique with the problems of reliability it implies [6, 51].
Unlike them, our solution is based on a more reliable labeling technique [50, 100]
and is evaluated with a realistic dataset for evolving data streams (i.e., 13 years
5.7. CHAPTER SUMMARY
97
of traffic, 4 000 millions of flows). We also perform a complete study of HAT in
order to understand the impact of its different parameters on the classification of
network traffic.
The problems that arise when a technique is deployed in an actual scenario
have been scarcely studied in the literature. To the best of our knowledge, only
our previous work [61] have addressed the problem of automatically updating the
classification models without human intervention. However, the features of this
new proposal considerably reduces the requirements of [61]. Although it also needs
a small sample of labeled traffic to keep the model updated, no data is stored nor
periodically retraining is performed. These appealing features makes our proposal
a solution very easy to maintain and deploy.
5.7
Chapter Summary
In this chapter, we proposed a new stream-based classification solution based on
Hoeffding Adaptive Trees. This technique has very appealing features for network
traffic classification: (i) processes an instance at a time and inspects it only once,
(ii) uses a predefined amount of memory, (iii) works in a bounded amount of
time and (iv) is ready to predict at any time. Furthermore, our technique is able
to automatically adapt to the changes of the traffic with just a small sample of
labeled data, making our solution very easy to maintain. As a result, we are able
to accurately classify the traffic using only NetFlow v5 data, which is already
provided by most routers at no cost, making our solution very easy to deploy.
We evaluate our technique using the publicly available MAWI dataset, 4 000
millions of flows from 15-minutes traces daily collected in a transit link in Japan
since 2001 (13 years). We first evaluate the impact of the different parameters on
the HAT technique when used for traffic classification and then compare it with
one of the state-of-the-art techniques most commonly used in the literature (i.e.,
C4.5).
The results show that our technique is an excellent solution for network traffic
classification. It is not only more accurate than traditional batch-based techniques, but it also sustains this very high accuracy over the years with less cost.
Furthermore, our technique does not require complex, ad-hoc retraining systems
to keep the system updated, which facilitates its deployment and maintenance in
operational networks.
98
CHAPTER 5. STREAMING-BASED TRAFFIC CLASSIFICATION
Chapter 6
Validation of Traffic Classification
Methods
6.1
Introduction
As described in Chapter 1, most traffic classification solutions proposed in the
literature report very high accuracy. However, these solutions mostly base their
results on a private ground-truth (i.e., dataset), usually labeled by techniques of
unknown reliability (e.g., ports-based or DPI-based techniques [46, 48–50]). That
makes it very difficult to compare and validate the different proposals. The use
of private datasets is derived from the lack of publicly available datasets with
payload. Mainly because of privacy issues, researchers and practitioners are not
allowed to share their datasets with the research community.
Another crucial problem is the reliability of the techniques used to set the
ground-truth. Most papers show that researchers usually obtain their groundtruth through port-based or DPI-based techniques [46,48–50]. The poor reliability
of port-based techniques is already well known, given the use of dynamic ports or
well-known ports of other applications [1, 51]. Although the reliability of DPIbased techniques is still unknown, according to conventional wisdom they are, in
principle, one of the most accurate techniques.
This chapter presents two main contributions. First, we publish a reliable
labeled dataset with full packet payloads [68]. The dataset has been artificially
built in order to allow us its publication. However, we have manually simulated
different behaviors to make it as representative as possible. We used VBS [60] to
guarantee the reliability of the labeling process. This tool, further described in
99
100CHAPTER 6. VALIDATION OF TRAFFIC CLASSIFICATION METHODS
Table 6.1: DPI-based techniques evaluated
Name
PACE
OpenDPI
nDPI
L7-filter
Libprotoident
NBAR
Version
1.41 (June 2012)
1.3.0 (June 2011)
rev. 6391 (March 2013)
2009.05.28 (May 2009)
2.0.6 (Nov 2012)
15.2(4)M2 (Nov 2012)
Applications
1 000
100
170
110
250
85
Section 6.2.1, can label the flows with the name of the process that created them.
This allowed us to carefully create a reliable ground-truth that can be used as a
reference benchmark for the research community. Second, using this dataset, we
evaluated the performance and compared the results of 6 well-known DPI-based
techniques, presented in Table 6.1, which are widely used for the ground-truth
generation in the traffic classification literature.
These contributions pretend to be a first step towards the impartial validation
of network traffic classifiers. They also provide to the research community some
insights about the reliability of different DPI-based techniques commonly used in
the literature for ground-truth generation.
The remainder of this work is organized as follows. Section 6.2 describes the
methodology used to create and label the dataset. Section 6.3 compares the performance of the selected DPI-based techniques. Section 6.4 extracts the outcomes
from the results previously obtained. Section 6.5 briefly reviews the related work.
Finally, Section 6.6 concludes the work and outlines our future work.
6.2
Methodology
This section describes the methodology used to build and reliable label the dataset
used in this evaluation. First, the testbed used and the selected applications
studied are presented. Then, the extraction and classification processes to create
the dataset are detailed. Finally, the characteristics of the resulting dataset are
described.
6.2. METHODOLOGY
6.2.1
101
The Testbed
Our testbed is based on VMWare virtual machines (VM). We installed three VM
for our data generating stations and we equipped them with Windows 7 (W7),
Windows XP (XP), and Ubuntu 12.04 (LX). Additionally, we installed a server
VM for data storage.
As described in Section 2.3.4, to collect and accurately label the flows, we
adapted Volunteer-Based System (VBS) developed at Aalborg University [60]. The
task of VBS is to collect information about Internet traffic flows (i.e., start time of
the flow, number of packets contained by the flow, local and remote IP addresses,
local and remote ports, transport layer protocol) together with detailed information about each packet (i.e., direction, size, TCP flags, and relative timestamp to
the previous packet in the flow). For each flow, the system also collects the process
name associated with that flow. The process name is obtained from the system
sockets. This way, we can ensure the application associated to a particular traffic. Additionally, the system collects some information about the HTTP content
type (e.g., text/html, video/x-flv ). The captured information is transmitted to the
VBS server, which stores the data in a MySQL database. The design of VBS was
initially described in [60]. On every data generating VM, we installed a modified
version of VBS. The source code of the modified version was published in [101]
under a GPL license. The modified version of the VBS client captures full Ethernet frames for each packet, extracts HTTP URL and Referer fields. We added a
module called pcapBuilder, which is responsible for dumping the packets from the
database to PCAP files. At the same time, INFO files are generated to provide
detailed information about each flow, which allows us to assign each packet from
the PCAP file to an individual flow. We also added a module called logAnalyzer,
which is responsible for analyzing the logs generated by the different DPI tools,
and assigning the results of the classification to the flows stored in the database.
6.2.2
Selection of the Data
The process of building a representative dataset, which characterizes a typical user
behavior, is a challenging task, crucial on testing and comparing different traffic
classifiers. Therefore, to ensure the proper diversity and amount of the included
data, we decided to combine the data on a multidimensional level. Based on
w3schools statistics, we selected Windows 7 (55.3 % of all users), Windows XP
(19.9 %), and Linux (4.8 %) - state for January 2013. Apple computers (9.3 % of
overall traffic) and mobile devices (2.2 %) were left as future work. The selected
102CHAPTER 6. VALIDATION OF TRAFFIC CLASSIFICATION METHODS
applications are shown below.
• Web browsers: based on w3schools statistics: Chrome and Firefox (W7, XP,
LX), Internet Explorer (W7, XP).
• BitTorrent clients: based on CNET ranking: uTorrent and Bittorrent (W7,
XP), Frostwire and Vuze (W7, XP, LX)
• eDonkey clients: based on CNET ranking: eMule (W7, XP), aMule (LX)
• FTP clients: based on CNET ranking: FileZilla (W7, XP, LX), SmartFTP
Client (W7, XP), CuteFTP (W7, XP), WinSCP (W7, XP)
• Remote Desktop servers: built-in (W7, XP), xrdp (LX)
• SSH servers: sshd (LX)
• Background traffic: DNS and NTP (W7, XP, LX), NETBIOS (W7, XP)
The list of visited websites was based on the top 500 websites according to
Alexa statistics. We chose several of them taking into account their rank and the
nature of the website (e.g., search engines, social medias, national portals, video
websites) to assure the variety of produced traffic. These websites include: Google,
Facebook, YouTube, Yahoo!, Wikipedia, Java, and Justin.tv. For most websites we
performed several random clicks to linked external websites, which should better
characterize the real behavior of the real users and include also other websites
not included in the top 500 ranking. This also concerns search engines, from
which we manually generated random clicks to the destination web sites. Each
of the chosen websites was processed by each browser. In case it was required to
log into the website, we created fake accounts. In order to make the dataset as
representative as possible we have simulated different human behaviors when using
these websites. For instance, on Facebook, we log in, interact with friends (e.g.,
chat, send messages, write in their walls), upload pictures, create events or play
games. On YouTube, we watched the 10 most popular videos, which we randomly
paused, resumed, and rewound backward and forward. Also, we randomly made
some comments and clicked Like or Not like buttons. The detailed description
of actions performed with the services is listed in our technical report [100]. We
tested the P2P (BitTorrent and eDonkey) clients by downloading files of different
sizes and then leaving the files to be seeded for some time, in order to obtain
enough of traffic in both directions. We tried to test every FTP client using both
the active transfer mode (PORT) and passive transfer mode (PASV), if the client
supports such mode.
6.2. METHODOLOGY
6.2.3
103
Extracting the Data for Processing
Each DPI tool can have different requirements and features, so the extracting
tool must handle all these issues. The PCAP files provided to PACE, OpenDPI,
L7-filter, nDPI, and Libprotoident are accompanied by INFO files, which contain
the information about the start and end of each flow, together with the flow
identifier. Because of that, the software, which uses the DPI libraries, can create
and terminate the flows appropriately, as well as to provide the classification results
together with the flow identifier. Preparing the data for NBAR classification is
more complicated. There are no separate INFO files describing the flows, since the
classification is made directly on the router. We needed to extract the packets in
a way that allows the router to process and correctly group them into flows. We
achieved that by changing both the source and destination MAC addresses during
the extraction process. The destination MAC address of every packet must match
up with the MAC address of the interface of the router, because the router cannot
process any packet which is not directed to its interface on the MAC layer. The
source MAC address was set up to contain the identifier of the flow to which it
belongs, so the flows were recognized by the router according to our demands. To
the best of our knowledge, this is the first work to present a scientific performance
evaluation of NBAR.
6.2.4
The Classification Process
We designed a tool, called dpi benchmark, which can read the PCAP files and
provide the packets one-by-one to PACE, OpenDPI, L7-filter, nDPI, and Libprotoident. All the flows are started and terminated based on the information from
the INFO files. After the last packet of the flow is sent to the classifier, the tool
obtains the classification label associated with that flow. The labels are written to
the log files together with the flow identifier, which makes us later able to relate
the classification results to the original flows in the database. A brief description
of the DPI tools used in this study is presented in Table 6.1. Although some of
the evaluated tools have multiple configuration parameters, we have used in our
evaluation the default configuration for most of them. A detailed description of
the evaluated DPI tools and their configurations can be found in [100].
Classification by NBAR required us to set up a full working environment. We
used GNS3 - a graphical framework, which uses Dynamips to emulate our Cisco
hardware. We emulated the 7200 platform, since only for this platform supported
by GNS3 was available the newest version of Cisco IOS (version 15), which contains
104CHAPTER 6. VALIDATION OF TRAFFIC CLASSIFICATION METHODS
Flexible NetFlow. The router was configured by us to use Flexible NetFlow with
NBAR on the created interface. Flexible NetFlow was set up to create the flows
taking into account the same parameters as are used to create the flow by VBS.
On the computer, we used tcpreplay to replay the PCAP files to the router with
the maximal speed, which did not cause packet loss. At the same time, we used
nfacctd, which is a part of PMACCT tools, to capture the Flexible NetFlow records
sent by the router to the computer. The records, which contain the flow identifier
(encoded as source MAC address) and the name of the application recognized by
NBAR, were saved into text log files. This process is broadly elaborated in our
technical report [100].
6.2.5
The PAM Dataset
Our dataset contains 1 262 022 flows captured during 66 days, between February
25, 2013 and May 1, 2013, which account for 35.69 GB of pure packet data. Section 2.3.4 describes in detail the properties of this dataset. The classes together
with the number of flows and the data volume are shown in Table 2.6.
We have published this labeled dataset with full packet payloads in [68]. Therefore, it can be used by the research community as a reference benchmark for the
validation and comparison of network traffic classifiers.
6.3
Performance Comparison
This section provides a detailed insight into the classification results of different
types of traffic by each of the classifiers. All these results are summarized in
Table 6.2, where the ratio of correctly classified flows (i.e., precision or true positives), incorrectly classified flows (i.e., errors or false positives) and unclassified
flows (i.e., unknowns) are respectively presented. The complete confusion matrix
can be found in our technical report [100].
Regarding the classification of P2P traffic, eDonkey is the first application studied. Only PACE, and especially Libprotoident, can properly classify it (precision
over 94 %). nDPI and OpenDPI (that use the same pattern), as well as NBAR, can
classify almost no eDonkey traffic (precision below 1 %). L7-filter classifies 1/3 of
the flows, but it also produces many false positives by classifying more than 13 %
of the flows as Skype, NTP, and finger. The wrongly classified flows in nDPI were
labeled as Skype, RTP and RTCP, and in NBAR as Skype. The classification of
BitTorrent traffic, the second P2P application studied, is not completely achieved
6.3. PERFORMANCE COMPARISON
105
by any of the classifiers. PACE and Libprotoident achieve again the highest precision (over 77 %). The rest of the classifiers present severe problems to identify
this type of traffic. When misclassified, the BitTorrent traffic is usually classified
as Skype.
The performance of most DPI tools with more traditional applications is significantly higher. FTP traffic is usually correctly classified. Only L7-filter and NBAR
present problems to label it. The false positives produced by L7-filter are because
the traffic is classified as SOCKS. Table 6.2 also shows that all the classifiers can
properly classify DNS traffic. Similar results are obtained for NTP, which almost
all the classifiers can correctly classify it. However, NBAR completely miss the
classification of this traffic. SSH was evaluated in its Linux version. Table 6.2
shows that NBAR almost classified all the flows while the rest of classifiers labeled
more than 95 % of them.
Similar performance is also obtained with RDP, usually employed by VoIP
applications, as shown in Table 6.2. Again, L7-filter and NBAR can not classify
this application at all. The false positives for L7-filter, Libprotoident, and NBAR
are mainly due to Skype, RTMP, and H323, respectively.
Unlike previous applications, the results for NETBIOS are quite different. Surprisingly, NBAR and nDPI are the only classifiers that correctly label NETBIOS
traffic. PACE can classify 2/3 of this traffic and OpenDPI only 1/4. On the
other hand, the patterns from L7-filter and Libprotoident do not properly detect
this traffic. The wrongly classified flows in Libprotoident are labeled as RTP and
Skype, and in L7-filter as eDonkey, NTP, and RTP.
We also evaluated RTMP traffic, a common protocol used by browsers and
plugins for playing FLASH content. It is important to note that only Libprotoident
has a specific pattern for RTMP. Because of that, we have also counted as correct
the RTMP traffic classified as FLASH although that classification is not as precise
as the one obtained by Libprotoident. L7-filter and NBAR can not classify this
type of traffic. The rest of the classifiers achieve a similar precision, around 80 %.
The surprising amount of false positives by nDPI is because some traffic is classified
as H323. L7-filter errors are due to wrongly classified traffic as Skype and TSP.
Table 6.2 also presents the results regarding the HTTP protocol. All of them
but L7-filter can properly classify most of the HTTP traffic. L7-filter labels all the
traffic as finger or Skype. nDPI classifies some HTTP traffic as iMessage Facetime.
The amount of errors from PACE is surprising, as this tool is usually characterized
by very low false positive ratio. All the wrong classifications are labeled as Meebo
traffic. The older Meebo pattern available in OpenDPI and the newer from nDPI
seems not to have this problem.
106CHAPTER 6. VALIDATION OF TRAFFIC CLASSIFICATION METHODS
Most incorrect classifications for all the tools are due to patterns that easily
match random traffic. This problem especially affects L7-filter and, in particular,
with the patterns used to match Skype, finger and NTP traffic. The deactivation
of those patterns would considerably decrease the false positive ratio but it would
disable the classification of those applications. In [46], the authors use a tailormade configuration and post-processing of the L7-filter output in order to minimize
this overmatching problem.
6.3.1
Sub-classification of HTTP Traffic
Our dataset also allows the study of HTTP traffic at different granularity (e.g.,
identify different services running over HTTP). However, only nDPI can subclassify some applications at this granularity (e.g., Youtube, Facebook ). Newer
versions of PACE also provide this feature but we had no access to it for this
study. Table 6.3 presents the results for four applications running over HTTP
identified by nDPI. Unlike the rest of tools that basically classify this traffic as
HTTP, nDPI can correctly give the specific label with precision higher than 97 %.
Furthermore, the classification errors are caused by traffic that nDPI classifies as
HTTP without providing the lower level label.
Another sub-classification that can be studied with our dataset is the FLASH
traffic over HTTP. However, the classification of this application is different for
each tool making its comparison very difficult. PACE, OpenDPI and nDPI have
a specific pattern for this application. At the same time, these tools (as well as
L7-filter) have specific patterns for video traffic, which may or may not run over
HTTP. In addition, nDPI has specific labels for Google, Youtube and Facebook
that can also carry FLASH traffic. Libprotoident and NBAR do not provide any
pattern to classify FLASH traffic over HTTP. Table 6.4 shows that nDPI can
correctly classify 99.48 % of this traffic, 25.48 % of which is classified as Google,
Youtube or Facebook. PACE and OpenDPI can properly classify around 86 % of
the traffic. The errors produced in the classification are almost always related to
traffic classified as HTTP with the exception of L7-filter that classifies 86.49 % of
the traffic as finger.
6.4
Lessons Learned and Limitations
This section extracts the outcomes from the results obtained during the performance comparison. Also, we discuss the limitations of our study. Table 6.5 presents
6.4. LESSONS LEARNED AND LIMITATIONS
107
Table 6.2: DPI evaluation by application
Application
eDonkey
BitTorrent
FTP
DNS
NTP
SSH
RDP
NETBIOS
RTMP
HTTP
Classifier
% correct
% wrong
% uncl.
PACE
OpenDPI
L7-filter
nDPI
Libprotoident
NBAR
PACE
OpenDPI
L7-filter
nDPI
Libprotoident
NBAR
PACE
OpenDPI
L7-filter
nDPI
Libprotoident
NBAR
PACE
OpenDPI
L7-filter
nDPI
Libprotoident
NBAR
PACE
OpenDPI
L7-filter
nDPI
Libprotoident
NBAR
PACE
OpenDPI
L7-filter
nDPI
Libprotoident
NBAR
PACE
OpenDPI
L7-filter
nDPI
Libprotoident
NBAR
PACE
OpenDPI
L7-filter
nDPI
Libprotoident
NBAR
PACE
OpenDPI
L7-filter
nDPI
Libprotoident
NBAR
PACE
OpenDPI
L7-filter
nDPI
Libprotoident
NBAR
94.80
0.45
34.21
0.45
98.39
0.38
81.44
27.23
42.17
56.00
77.24
27.44
95.92
96.15
6.11
95.69
95.58
40.59
99.97
99.97
98.95
99.88
99.97
99.97
100.00
100.00
99.83
100.00
100.00
0.40
95.57
95.59
95.71
95.59
95.71
99.24
99.04
99.07
0.00
99.05
98.83
0.00
66.66
24.63
0.00
100.00
0.00
100.00
80.56
82.44
0.00
78.92
77.28
0.23
96.16
98.01
4.31
99.18
98.66
99.58
0.02
0.00
13.70
6.72
0.00
10.81
0.01
0.00
8.78
0.43
0.06
1.49
0.00
0.00
93.31
0.45
0.00
0.00
0.00
0.00
0.13
0.09
0.00
0.02
0.00
0.00
0.15
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.05
0.02
0.02
91.21
0.08
0.16
0.66
0.08
0.00
8.45
0.00
5.03
0.00
0.00
0.00
24.12
8.90
0.47
0.23
1.85
0.00
95.67
0.76
0.00
0.00
5.18
99.55
52.09
92.83
1.60
88.81
18.54
72.77
49.05
43.58
22.71
71.07
4.08
3.85
0.57
3.85
4.42
59.41
0.03
0.03
0.92
0.03
0.04
0.02
0.00
0.00
0.02
0.00
0.00
99.60
4.43
4.41
4.29
4.41
4.30
0.70
0.94
0.91
8.79
0.87
1.01
99.34
33.26
75.37
91.55
0.00
94.97
0.00
19.44
17.56
75.88
12.18
22.25
99.53
1.99
1.99
0.02
0.06
1.34
0.42
the summary of the results from Section 6.3. The Precision (i.e., first column) is
computed similarly to Section 6.3, but we take into account all the applications
together (i.e., 100 * # correctly classified flows / # total flows). However, this
metric is dependent on the distribution of the dataset. Because of that, we also
108CHAPTER 6. VALIDATION OF TRAFFIC CLASSIFICATION METHODS
Table 6.3: HTTP sub-classification by nDPI
Application
Google
Facebook
Youtube
Twitter
% correct
97.28
100.00
98.65
99.75
% wrong
2.72
0.00
0.45
0.00
% unclassified
0.00
0.00
0.90
0.25
Table 6.4: FLASH evaluation
Classifier
PACE
OpenDPI
L7-filter
nDPI
Libprotoident
NBAR
% correct
86.27
86.34
0.07
99.48
0.00
0.00
% wrong
13.18
13.15
99.67
0.26
98.07
100.00
% unclassified
0.55
0.51
0.26
0.26
1.93
0.00
compute a second metric, the Average Precision. This statistic is independent
from the distribution and is calculated as follow:
PN correctly classif ied i f lows
Avg. P recision =
i=1
total i f lows
N
(6.1)
where N is the number of applications studied (i.e., N = 10).
As it can be seen in Table 6.5, PACE is the best classifier. Even while we were
not using the last version of the software, PACE was able to properly classify 94 %
of our dataset. Although Ma et al. already pointed out in [102] the importance of
the first bytes of payload for the classification, surprisingly for us, Libprotoident
achieves similar results only inspecting the first four bytes of payload for each
direction. On the other hand, L7-filter and NBAR perform poorly in classifying
the traffic from our dataset. The more fair metric, Avg. Precision, presents similar
results. PACE is still the best classifier, however, it has increased the difference by
several points to the second best classifier, Libprotoident. Unlike before, nDPI is
almost as precise as Libprotoident with this metric. L7-filter and NBAR are still
the tools that present the worst performance.
Nonetheless, the previous conclusions are obviously tied to our dataset. Although we have tried our best to emulate the real behavior of the users, many
6.5. RELATED WORK
109
Table 6.5: DPI summary results
Classifier
PACE
OpenDPI
L7-filter
nDPI
Libprotoident
NBAR
% Precision
94.22
52.67
30.26
57.91
93.86
21.79
% Avg. Precision
91.01
72.35
38.13
82.48
84.16
46.72
applications, behaviors and configurations are not represented on it. Because of
that, it has some limitations. In our study we have evaluated 10 well-known
applications, however adding more applications as Skype or Spotify is part of our
ongoing future work. The results obtained from the different classifiers are directly
related to those applications. Thus, the introduction of different applications could
arise different outcomes. The traffic generated for building the dataset, although
has been manually and realistically created, is artificial. The backbone traffic
would carry different behaviors of the applications that are not fully represented
in our dataset (e.g., P2P clients running on port 80). Therefore, the performance
of the tools studied could not be directly extrapolated from the current results,
but it gives an idea of their precision in the evaluated set of applications. At the
same time, the artificially created traffic allowed us to publish the dataset with
full packet payloads.
6.5
Related Work
Some previous works evaluated the accuracy of DPI-based techniques [49, 50,
103, 104]. These studies rely on a ground-truth generated by another DPI-based
tool [50], port-based technique [49] or a methodology of unknown reliability [103,
104], making their comparison very difficult. Recently, a concomitant study to
ours [104] compared the performance of four DPI-based techniques (i.e., L7-filter,
Tstat, nDPI and Libprotoident). This parallel study confirms some of the findings
of our work presenting nDPI and Libprotoident as the most accurate open-source
DPI-based techniques. In [105] the reliability of L7-filter and a port-based technique was compared using a dataset obtained by GT [106] showing that both
techniques present severe problems to accurately classify the traffic.
To the best of our knowledge, just one work has tackled the problem of the lack
110CHAPTER 6. VALIDATION OF TRAFFIC CLASSIFICATION METHODS
of publicly available labeled datasets. Gringoli et al. in [106] published anonymized
traces without payload, but accurately labeled using GT. This dataset is very
interesting to evaluate Machine Learning-based classifiers, but the lack of payload
makes it unsuitable for DPI-based evaluation.
6.6
Chapter Summary
This work presents the first step towards validating the reliability of the accuracy
of the network traffic classifiers. We have compared the performance of six tools
(i.e., PACE, OpenDPI, L7-filter, nDPI, Libprotoident, and NBAR), which are
usually used for the traffic classification. The results obtained in Section 6.3 and
further discussed in Section 6.4 show that PACE is, on our dataset, the most
reliable solution for traffic classification. Among the open-source tools, nDPI and
especially Libprotoident present the best results. On the other hand, NBAR and
L7-filter present several inaccuracies that make them not recommendable as a
ground-truth generator.
In order to make the study trustworthy, we have created a dataset using
VBS [60]. This tool associates the name of the process to each flow making its
labeling totally reliable. The dataset of more than 500 K flows contains traffic
from popular applications like HTTP, eDonkey, BitTorrent, FTP, DNS, NTP,
RDP, NETBIOS, SSH, and RDP. The total amount of data properly labeled is
32.61 GB. Furthermore, and more important, we release to the research community this dataset with full payload, so it can be used as a common reference for
the comparison and validation of network traffic classifiers.
Chapter 7
Other Improvements to Traffic
Classifiers
This chapter collects other contributions achieved in parallel with the main theme
of this thesis. Section 7.1 presents the study of an efficient profiled flow termination
timeout. Although this study is directly related to network traffic classification
field, its application is not limited to this topic. Most monitoring tools continuously
keep the state of the connections (i.e., flows) of the network monitored. An efficient
implementation of this process is crucial in scenarios with limited resources and
huge amount of connections. As a result, a proper mechanism for expiration of the
flows is a key point in the optimization of this costly process. Section 7.1 studies
a mechanism for flow termination by group of applications with current Internet
traffic. From that characterization we propose an expiration technique to achieve
optimized timeouts in a profiled way.
Section 7.2 address the classification problem from a different point of view.
The last contribution of this thesis proposes an early classification technique able
to classify the traffic just using the size of the first packets of the flows. This early
classification feature is very important, since it allows network operators to quickly
react to the classification results (e.g., traffic shaping). In addition, we present the
prototype of the system for continuous network traffic classification described in
Chapter 4.
111
112
CHAPTER 7. OTHER IMPROVEMENTS TO TRAFFIC CLASSIFIERS
7.1
7.1.1
Efficient Profiled Flow Termination Timeouts
Introduction
The increase of bandwidth and diversity of protocols on the Internet has made
the online traffic classification a highly difficult and exigent task. In current networks (e.g. 10/40/100 Gbps), packets are received every few ns, and that is the
time available for parsing them without requiring data buffering. However, many
proposed traffic classifiers require to store statistics or even the actual content of
the packets for the classification at flow level. Furthermore, it is usually necessary
to keep the state of each flow in order to properly classify it. Because of this, the
operations involved in this process have to be carried out very efficiently in terms
of memory and CPU.
In this section, we focus our attention in the optimization of the memory consumption. A crucial issue in this process is the expiration of the structure that
keeps the state of each flow. On the one hand, expiring the flows too late incredibly increase the amount of memory necessary. On the other hand, expiring the
flows very quickly would create segmented flows with less information, making the
accuracy of classification more difficult or even impossible. As a result, a proper
expiration of the flows is a key point in the optimization of this costly process. So
far, this process has been usually carried out using timeouts or specific protocol
mechanisms (e.g., TCP termination handshake).
The first contribution of this work is a comprehensive study of the flow termination by group of applications with current Internet traffic. The study shows
that the non-desired flow terminations (e.g., RST, undefined termination) are considerably common in the current traffic. Furthermore, we have detected different
behaviors related to the flow termination among the different groups of applications. From that characterization, we have gone further and proposed an expiration
technique to achieve optimized timeouts in a profiled way. The proposed method
has been evaluated using the PACE engine [10] achieving a substantial reduction
of memory while maintaining the accuracy and the CPU consumption.
The rest of the work is organized as follows. Section 7.1.2 describes the traces
used. Section 7.1.3 briefly presents the methodology used to extract the results.
Section 7.1.4 presents the statistical results, and based on them, the estimation
and evaluation of the timeouts is done. Section 7.1.5 reviews the related work.
Finally, Section 7.1.6 concludes the study.
7.1. EFFICIENT PROFILED FLOW TERMINATION TIMEOUTS
7.1.2
113
Dataset
The traces used for the evaluation were recorded in two different network scenarios
and they provide a good comparative basis for the results. These traces were
taken under confidential agreements, which means that specific details can not be
revealed. In Table 7.1 are synthesized some of their properties.
ISP Core This trace was recorded in 2009 in an internal link of a Tier-1 ISP,
after the access and aggregation sections and before the output router to the
backbone. This type of networks are often designed with multiple links due to
balancing and failure issues. As a consequence of that, many of the flows recorded
are asymmetric.
ISP Mob This trace is a good complement for the ISP Core one since it was
recorded in a mobile operator network. It was taken somewhere between the SGSN
and the GGSN, elements of a standard GPRS network. As it is near to the edge,
this trace contains a high proportion of symmetric traffic. The transmission of the
data is done over GTP, however, PACE includes features for decapsulation. It was
recorded in 2010.
Table 7.1: Characteristics of the ISP traces in the evaluation dataset
ISP Core
ISP Mob
7.1.3
Duration
38 160 s
2 700 s
Flows
295 729 K
6 093 K
Packets
7 074 618 K
233 359 K
Bytes
2 591 636 M
133 046 M
Methodology
The main tool employed in this study is based on the ipoque’s PACE engine [10],
the commercial DPI library. Thus, the classification in several protocol groups
has the reliability of this product (i.e., they claim nearly 100% detection rate of
protocols and applications with no false positive). The protocol groups considered
are the following: Generic, P2P, Gaming, Tunnel, VoIP, IM, Streaming, Mail Network management, Filetransfer, and Web. The Generic group is mainly composed
by unknown traffic.
In order to derive the set of timeouts for the optimization of the expiration
process, we have developed a module for PACE that calculates different statistics by group of applications. Although this module computes different statistics
114
CHAPTER 7. OTHER IMPROVEMENTS TO TRAFFIC CLASSIFIERS
for each flow, the most important for our purpose are the times between packets,
named PCK-PCK, and the times between packets with data, named DAT-DAT.
The difference among them is that the PCK-PCK times include the TCP’s acknowledgment packets (ACK) while the DAT-DAT times include only the packets
with application data. All the evaluations are done both globally and considering
each of the Client-Server and Server-Client direction. In addition, we also study
the termination processes of the different group of applications for the TCP traffic. The terminations detected are the standard FIN process, the RST process, the
RST process in a blocking scenario, and the undefined. These termination modes
are further described in Section 7.1.4.
To avoid, as far as possible, the inclusion of worthless data from our traces,
the following considerations are taken. For TCP traffic, only those flows showing
the first step of the 3 way-handshake (i.e., the SYN-SYNACK process) are parsed.
Doing so, we avoid considering ongoing flows initiated before the beginning of the
traces and we only consider bidirectional flows. On the other hand, due to its
intrinsic characteristics, for UDP traffic the only condition is to see at least two
packets (otherwise any PCK-PCK time could not be calculated). In Table 7.2 can
be seen the impact of applying these considerations on each trace.
Table 7.2: Flow usage after the sanitization process
ISP Core
ISP Mob
TCP
159 444 K
3 904 K
UDP
127 930 K
2 063 K
TCP used
42 521 K
3 454 K
UDP used
55 182 K
1 850 K
The establishment of Client-Server or Server-Client direction is easy for TCP
traffic, the side that starts the connection (i.e., SYN message) is considered as the
client. On the other hand, for UDP traffic the consideration is not so straightforward. First, it is checked if PACE can determine the direction based on the
protocol information. If it is not able, or the direction is either Client-Client or
Server-Server, we assume that the first packet seen comes from the client.
7.1.4
Results
The results obtained in this study are presented in two parts. First, the results
observed for the PCK-PCK and DAT-DAT times and the proportion of the different types of flow termination are discussed. Second, using these previous results,
a set of timeouts by application group are derived for an efficient flow expiration.
7.1. EFFICIENT PROFILED FLOW TERMINATION TIMEOUTS
115
Additionally, we show in Table 7.3 the traffic distribution of the traces used in
the evaluation. The percentages of each trace can be seen for both the TCP and
UDP traffic. The heterogeneous composition of these two traces from completely
different scenarios supports the representativeness of the results of our study. Since
we have studied the already cited protocol groups (see Section 7.1.3), the excess
traffic is grouped as Other.
Table 7.3: Traffic mix by flow in the evaluation dataset
Protocol
Generic
P2P
Gaming
Tunnel
VoIP
IM
Streaming
Mail
Management
Filetransfer
Web
Other
TCP Core
12.57%
5.98%
0.02%
10.66%
0.84%
10.60%
1.07%
14.17%
0.01%
0.69%
42.98%
0.41%
TCP Mob
10.02%
2.57%
0.00%
9.29%
0.07%
0.46%
0.71%
1.50%
0.01%
0.30%
74.69%
0.38%
UDP Core
14.16%
13.81%
0.03%
0.02%
1.94%
0.35%
58.33%
0.00%
11.35%
0.00%
0.00%
0.01%
UDP Mob
13.22%
13.21%
0.08%
0.30%
1.05%
0.05%
0.42%
0.00%
69.93%
0.00%
0.00%
1.74%
Time profiled study
Inter-packet times evaluation This section presents the results related to the
PCK-PCK and DAT-DAT times (see Section 7.1.3) measured for both the TCP
and UDP flows. They will be fundamental to establish the flow termination timeouts, especially for UDP traffic, as further explained in Section 7.1.4. Additionally,
some interesting outcomes are extracted from the statistics obtained.
• Similar PCK-PCK times over TCP regardless of the scenario: Excluding the Network management group, there is a strong correlation comparing the PCK-PCK values between the ISP Core (see Table 7.4) and
ISP Mob (see Table 7.5). This is also accomplished for the DAT-DAT values,
but not for the UDP traffic (which is always DAT-DAT). This result suggests
that the average packet cadence over TCP is independent on the network.
116
CHAPTER 7. OTHER IMPROVEMENTS TO TRAFFIC CLASSIFIERS
• PCK-PCK similarities in Client-Server and Server-Client: In both
scenarios and for the TCP (see Tables 7.4 and 7.5) and the UDP (see Table
7.6) traffic, there are similar PCK-PCK values comparing the Client-Server
and Server-Client directions. In the UDP case, it could be expected to see
an asymmetric behavior in some protocol groups like the Streaming as it
happens in the DAT-DAT values over TCP (see Table 7.7). This highlights
the different role that it takes over UDP.
• TCP and UDP usage differences: UDP flows for some protocol groups
in the ISP Core show (see Table 7.6) a lower or higher packet transmission
rate than the DAT-DAT seen over TCP (see Table 7.7), or as just explained
for the Streaming group it has an unexpected behavior. That is justified by
the usage of control flows over TCP or UDP by some protocols. For example,
this behavior is seen in the P2P, where the DAT-DAT times are around 4
times lower in the TCP case (probably real data transfer) than in the UDP
(probably control flows) one. This behavior also occurs but in the opposite
way for the Gaming, Tunnel, VoIP, and IM groups, which would imply that
in this case the real stream of data is sent over UDP while the TCP flows
are the control ones. This behavior is similar in the ISP Mob trace, with
the main exception seen for the Streaming group since in this case the UDP
traffic seems to carry as well real traffic (low value for the PCK-PCK times).
Table 7.4: PCK-PCK times TCP in ISP Core trace by application (ms)
Group
Generic
P2P
Gaming
Tunnel
VoIP
IM
Stream
Mail
Manage
Filetx
Web
Avg.
1 510
1 196
209
1 029
1 688
2 715
120
518
1 494
390
1 127
STD
12 860
7 084
1 319
11 080
9 981
13 356
3 200
8 259
12 774
36 037
13 383
Avg. c-s
2 760
2 216
448
2 004
3 161
5 240
286
802
2 430
855
2 294
STD c-s
17 840
9 201
1 777
15 412
13 677
17 857
4 842
10 435
17 382
55 946
18 757
Avg. s-c
2 720
2 217
384
1 737
3 450
5 302
181
1 040
3 313
647
1 827
STD s-c
16 740
9 593
1 732
14 706
14 137
18 390
3 871
11 959
19 552
46 807
17 586
7.1. EFFICIENT PROFILED FLOW TERMINATION TIMEOUTS
117
Table 7.5: PCK-PCK times TCP in ISP Mob trace by application (ms)
Group
Generic
P2P
Gaming
Tunnel
VoIP
IM
Stream
Mail
Manage
Filetx
Web
Avg.
1 796
1 907
109
1 396
1 976
1 897
87
1 057
316
427
704
STD
14 293
8 660
824
20 782
11 315
15 667
2 269
13 678
4 374
9 819
7 349
Avg. c-s
3 043
3 475
139
2 778
3 744
3 587
216
2 135
604
700
1 364
STD c-s
18 868
11 648
946
29 860
15 264
21 663
3 564
20 005
6 055
13 992
9 793
Avg. s-c
3 526
3 579
473
2 414
3 892
3 894
117
1 957
682
808
1 117
STD s-c
19 840
13 058
2 520
27 576
15 843
22 373
2 870
18 696
6 527
14 469
8 861
Table 7.6: PCK-PCK times UDP in ISP Core trace by application (ms)
Group
Generic
P2P
Gaming
Tunnel
VoIP
IM
Stream
Manage
Avg.
2 878
17 310
199
454
306
151
4 375
11 161
STD
23 551
64 100
2 997
7 215
7 888
1 471
37 134
55 806
Avg. c-s
3 427
26 099
295
572
509
209
7 068
20 342
STD c-s
25 744
77 177
3 816
8 284
10 211
1 829
47 118
73 220
Avg. s-c
4 719
18 017
374
1 061
404
198
5 846
21 947
STD s-c
27 850
70 828
3 948
12 091
10 430
3 171
43 562
79 760
Flow termination evaluation In order to obtain the set of profiled timeouts we
calculate not only the timing parameters but also the proportion of the different
flow termination types. This study is carried out for the TCP traffic given its
different termination mechanism. Four possible terminations are observed: TCP
handshake, RST, undefined termination, and RST in a blocking scenario. Fig. 7.1
shows the behavior of the termination of a RST in a blocking scenario. This is
when one or both extremes of the communication do not receive the corresponding
packets from the other side.
Unlike we expected, as we can see in Fig. 7.2, the number of flows not finished
by the standard TCP handshake is very high. Additionally, there is a big amount
118
CHAPTER 7. OTHER IMPROVEMENTS TO TRAFFIC CLASSIFIERS
Table 7.7: DAT-DAT times TCP in ISP Core trace by application (ms)
Group
Generic
P2P
Gaming
Tunnel
VoIP
IM
Stream
Mail
Manage
Filetx
Web
Avg.
2 137
1 772
372
1 334
2 656
4 321
136
746
1 278
533
880
STD
14 585
8 077
1 521
13 057
12 609
16 135
2 960
10 423
10 442
46 120
9 269
Avg. c-s
8 673
3 763
1 020
3 465
5 571
12 360
3 230
826
7 825
2 944
4 714
STD c-s
41 852
14 747
4 274
24 198
19 021
37 903
18 987
11 717
33 821
130 181
24 456
Avg. s-c
2 566
2 741
574
1 584
4 825
6 167
137
1 945
2 634
589
911
STD s-c
16 870
11 162
2 271
23 618
18 193
38 115
3 055
19 036
13 823
49 281
9 843
of flows terminating with a RST in a blocking scenario. In these situations, the
retransmission times grow exponentially [107, 108]. Thus, we have particularly
studied the times for the flows finished with a RST in a blocking scenario (i.e.,
time between a RST and the last packet of the flow, and PCK-PCK time). They
are shown in Table 7.8. Notice that these PCK-PCK times are generally bigger
than the standard ones (see Table 7.4). The results for the ISP Mob are very
similar and are omitted for the sake of space. These results can find a small
comparison basis in [109].
Estimation and evaluation of profiled timeouts
Once studied the times between packets and flow terminations by group of applications, we can proceed with the study of the profiled timeouts. A flow timeout
has to be big enough to not consider new flows with a late packet of an existing
one, but as low as possible to release the memory as soon as possible. An old
previous study already defined the use of a global timeout of 64 seconds. [110].
However, this value has lost its reliability because of the evolution of the different
applications on the Internet. In order to address this problem, we have found a
way to establish it according to different protocol groups by using the statistics
obtained, basically the PCK-PCK times. For TCP traffic, also are considered the
times regarding the PCK-PCK after a RST in a blocking scenario (see Table 7.8)
and the proportion of the different terminations. It is important since in these
7.1. EFFICIENT PROFILED FLOW TERMINATION TIMEOUTS
119
Figure 7.1: Blocking scenario behavior. Both sides send data, but since there is
no acknowledgment from the other side they try to retransmit (after a time which
grows in every attempt) until finally break the connection with a RST
Table 7.8: RST times in ISP Core trace by application (ms)
Group
Generic
P2P
Gaming
Tunnel
VoIP
IM
Stream
Mail
Manage
Filetx
Web
Avg. rst-end
34 276
15 095
57 834
13 618
20 311
8 819
7 449
29 384
182
11 967
11 534
STD rst-end
138 999
53 451
128 352
96 933
91 428
68 851
50 738
126 616
259
86 210
59 228
Avg. pck
6 722
6 354
765
5 618
4 753
1 964
344
9 357
157
3 768
1 996
STD pck
33 147
18 664
5 673
33 423
29 569
17 671
4 060
38 504
239
45 231
13 051
situations the time of inactivity grows exponentially and forces a high timeout.
The selection of a threshold for the timeouts is done considering the corresponding statistical results, and it is set up as {average + 2.5 ∗ ST D}. This formula
allows us to exclude unusual big measures, while still representing around 99% of
the complete time values. For each trace and separately for the TCP and UDP
120
CHAPTER 7. OTHER IMPROVEMENTS TO TRAFFIC CLASSIFIERS
Figure 7.2: Termination proportions in ISP Core trace, green stands for the standard FIN process, yellow for the unclosed flows, red for the RST and intense red
for the RST in a blocking scenario
flows, we choose the biggest threshold from either the Client-Server or the ServerClient direction. Thereby, we consider always the worst case, making it valid as
well when dealing with unidirectional flows. Thus, we define these thresholds as
t\
pck blk and the PCK-PCK times
pck pck for the PCK-PCK standard times and t\
after a RST in a blocking scenario.
We have seen that t\
pck blk times are forcing a higher value when considering the
optimal timeout. For that reason, we have used these values to round off the final
results, making them closer to the t\
pck pck times. The criteria used is based in two
aspects. On the one hand, we have considered which proportion of the traffic is
behaving as in a blocking scenario. On the other, we have also studied the CDF of
the number of packets in relation to the timeout for each group to see what is its
variability. That is, what is the effect of reducing the timeout in the proportions of
values inside the t\
pck blk interval. Fig. 7.3 shows the three groups we have detected.
A strong variation is seen for the P2P, VoIP, Streaming, and Network management
groups. A medium variation for the Generic, Tunnel, IM, Mail, and Web groups.
Finally, a slow variation is seen for the Gaming and Filetransfer groups.
7.1. EFFICIENT PROFILED FLOW TERMINATION TIMEOUTS
(a) Strong variation (P2P)
121
(b) Medium variation (Web)
(c) Slow variation (Filetransfer)
Figure 7.3: CDF of inter-packet times with RST termination in a blocking scenario
It is difficult to include these behaviors in the calculation of the final timeout.
Notwithstanding, we have done it applying a factor, named F , which modifies
the final timeout taking into account the behavior above described. This factor
is set up as {0.33, 0.66, 1} for the slow, medium and strong variation respectively.
Thus, we have considered that for TCP traffic a good theoretical timeout could be
approximated as follows:
ttimeout = (1 − f inblk ) ∗ t\
pck blk ∗ F
pck pck + f inblk ∗ t\
(7.1)
F = {0.33, 0.66, 1}
(7.2)
Remember the t\
pck blk regards the RST times in a blocking scenario. Therefore,
the proportion f inblk is as well related to that field, but concerning the proportion
of termination modes. Both parameters are considered getting the worst case value
among the ISP Core and the ISP Mob (i.e., the case where the timeout is bigger),
122
CHAPTER 7. OTHER IMPROVEMENTS TO TRAFFIC CLASSIFIERS
so we intend to find out a global timeout independent of the traces evaluated.
By applying this formula, we consider the t\
pck blk according to its proportion of
appearance in terms of flows termination and also the behavior seen from the
CDF. The parameters used for this calculation and the results are shown in Table
7.9. For UDP traffic, no rounding off has been done since the only parameter to
consider is the t\
pck pck .
Table 7.9: Evaluation of TCP timeout by application (ms)
Group
Generic
P2P
Gaming
Tunnel
VoIP
IM
Stream
Mail
Manage
Filetx
Web
t\
pck pck
53 126
36 255
6 772
77 426
43 500
59 826
12 392
52 147
52 192
140 720
47 804
t\
pck blk
117 269
78 086
26 440
119 298
90 671
163 225
35 098
127 211
76 480
497 736
83 657
f inblk
0.1326
0.181
0.1246
0.1244
0.1008
0.0787
0.3653
0.1011
0.1018
0.291
0.0428
F
0.66
1
0.33
0.66
1
0.66
1
0.66
1
0.33
0.66
ttimeout
56 344
43 826
7 015
77 589
48 255
63 596
20 687
55 363
54 665
147 568
48 121
Table 7.10: Evaluation of UDP timeout by application (ms)
Group
Generic
P2P
Gaming
Tunnel
VoIP
IM
Streaming
Management
ttimeout
74 343
219 043
45 584
40 814
71 636
14 030
124 862
221 346
In addition, we have gone further and we have evaluated the impact of these
outcomes by means of the PACE engine [10]. Thanks to that, we have been able to
measure not only the memory saving but also the detection rate impact. However,
7.1. EFFICIENT PROFILED FLOW TERMINATION TIMEOUTS
123
due to technical limitations in PACE, we have not been able to set up each timeout
in a profiled way for each protocol group. Notwithstanding, we have elaborated a
method to see what is the effect of varying the timeout on each group. To that
purpose, we have measured for a set of intervals from 20 to 300 seconds, in steps
of 20 seconds, what are the number of flows detected by PACE according to each
protocol group. We have also added the values for 600 seconds to have a wide
timeout to compare with. Then, we have plotted (see Fig. 7.4) the proportion of
flows per timeout compared with the maximum number of flows seen per protocol
group (when the timeout is 20 seconds). Therefore, we can get a curve to see when
the number of flows per protocol group tends to be stable.
It is only evaluated the ISP Core trace since it is more representative. However,
the results regarding the ISP Mob trace show similarities. It has to be considered
that for the FTP inside the Filetransfer group the timeout is directly established
by PACE and its operation is independent from the one we vary. That is why
in Fig. 7.4 it is not seen a wide variation for this group. Also, it is important
to comment the effect on the VoIP and lesser extent in the P2P group, since
at some point the variation increases, hindering the evaluation of our theoretical
timeouts. This happens due to internal detection methods of PACE. With the
exception of the Generic group (remember that it contains the unknown traffic)
and the already commented situations, and considering that the evaluation is done
over all the traffic tighter (i.e., TCP and UDP), the theoretical results are roughly
accomplished.
Similarly, we have also checked a global timeout by measuring what is the
direct impact in the memory requirements, the detection rate and the performance
of PACE. We can see in Fig. 7.5 that for a certain timeout both the detection
rate and the total number of flows (which is directly related to the performance)
tend to be stable. Comparing a timeout in which the detection rate and number
of flows stabilizes (approximately 200 seconds) with 600 seconds (this was the
reference timeout implemented by PACE), we achieve a reduction of around 60%
of memory necessities.
7.1.5
Related Work
To the best of our knowledge, this is the first work that presents a comprehensive
study about the flow termination by group of applications. Some papers in the
literature are related to this work but in a simplified way and considering just few
parameters. In [109], it is intended to find out what is the influence of different
types of traffic on the Internet in terms of volume of data and establishment and
124
CHAPTER 7. OTHER IMPROVEMENTS TO TRAFFIC CLASSIFIERS
Figure 7.4: Evaluation of PACE timeout in ISP Core trace by application
termination behavior. Notwithstanding, the classification only covers P2P and
HTTP traffic. Concerning the statistical study of different protocol groups, the
somehow related papers with the findings and objectives of this work are either old
or sharing just a few similarities, so they are not considered. The only exception
is [110]. Although this paper is quite old (1995), it establishes a single timeout
which can be compared to this work findings. The use of a single timeout for the
expiration of the flows is a mechanism commonly used by the traffic classification
techniques proposed in the literature [36, 46].
7.1.6
Section Summary
Nowadays Internet bandwidth has forced online traffic classifiers to be implemented
very efficiently. Among the different constraints that online traffic classification
7.1. EFFICIENT PROFILED FLOW TERMINATION TIMEOUTS
125
Figure 7.5: Performance of PACE timeout in ISP Core trace by flows
present, this section focuses on the optimization of the memory consumption. A
key issue in this aspect is the expiration of the structures that keep the state of
each flow. This work proposes a new methodology for an efficient expiration of
the flows based on profiled timeouts. In order to obtain these timeouts, has been
studied the behavior of the traffic by group of applications. Unexpected results
have been observed regarding the proportion and types of flow termination in the
current traffic.
The profiled timeouts obtained have been evaluated in a well-known commercial DPI tool (the ipoque’s PACE engine), achieving a drastic reduction of memory
while keeping the computation cost and classification accuracy. The results suggest that the implementation of the profiled timeouts in the traffic classification
techniques proposed in the literature [37] could considerably improve their memory
consumption, alleviating by this, their feasibility for online classification.
126
CHAPTER 7. OTHER IMPROVEMENTS TO TRAFFIC CLASSIFIERS
7.2
7.2.1
Early Classification of Network Traffic
Introduction
Gaining information about the applications that generate traffic in an operational
network is much more than mere curiosity for network operators. Traffic engineering, capacity planning, traffic management or even usage-based pricing are some
examples of network management tasks for which this knowledge is extremely
important. Although this problem is still far from a definitive solution, the networking research community has proposed several ML techniques for traffic classification that can achieve very promising results in terms of accuracy. However,
in practice, most network operators still use either obsolete (e.g., port-based) or
unpractical (e.g., pattern matching) methods for traffic identification and classification. One of the reasons that explains this slow adoption by network operators
is the time-consuming training phase involving most ML-based methods, which
often requires human supervision and manual inspection of network traffic flows.
In this work, we revisit the viability of using the well-known Nearest Neighbor
(NN) ML technique for traffic classification. As we will discuss throughout the
sections, this method has a large number of features that make it very appealing
for traffic classification. However, it is often discarded given its poor classification
speed [34, 36]. In order to address this practical problem, we present an efficient
implementation of the NN search algorithm based on a K-dimensional tree structure that allows us not only to classify network traffic online with high accuracy,
but also to retrain the classifier on-the-fly with minimum overhead, thus lowering
the barriers that hinder the general adoption of ML-based methods by network
operators.
Our K-dimensional tree implementation only requires information about the
length of the very first packets of a flow. This solution provides network operators
with the interesting feature of early classification [20,21]. That is, it allows them to
rapidly classify a flow without having to wait until its end, which is a requirement of
most previous traffic classification methods [16,26,30]. In order to further increase
the accuracy of our method along with its classification speed, we combine the
information about the packet sizes with the relevant data still provided by the
port numbers [36].
We present an actual implementation of our method based on the Traffic Identification Engine (TIE) [111]. TIE is a community-oriented tool for traffic classification that allows multiple classifiers (implemented as plugins) to run concurrently
and produce a combined classification result.
7.2. EARLY CLASSIFICATION OF NETWORK TRAFFIC
127
Given the low overhead imposed by the training phase of our method and
the plugins already provided by TIE to set the ground-truth (e.g., L7 plugin),
our implementation has the unique feature of continuous training. This feature
allows the system to automatically retrain itself as the training data becomes
obsolete. We hope that the large advantages of our method (i.e., accuracy (>
95%), classification speed, early classification and continuous training) can give
an incentive to network operators to progressively adopt new and more accurate
ML-based methods for traffic classification.
The remainder of this work is organized as follows. Section 7.2.2 describes
our ML-based method based on TIE. Section 7.2.3 analyzes the performance of
our method and presents preliminary results of its continuous training feature.
Section 7.2.4 reviews the related work. Finally, Section 7.2.5 concludes the work
and outlines our future work.
7.2.2
Methodology
This section describes our ML-based classification method based on multiple Kdimensional trees, together with its continuous training system. We also introduce
TIE, the traffic classification system we use to implement our technique, and the
modifications made to it in order to allow our method to continuously retrain itself.
Traffic Identification Engine
TIE [111] is a modular traffic classification engine developed by Universitá di
Napoli Federico II. This tool is designed to allow multiple classifiers (implemented
as plugins) to run concurrently and produce a combined classification result. In
this work, we implement our traffic classification method as a TIE plugin.
TIE is divided in independent modules that are in charge of the different classification tasks. The first module, P acket F ilter, uses the Libpcap library to collect
the network traffic. This module can also filter the packets according to BPF or
user-level filters (e.g., skip the first n packets, check header integrity or discard
packets in a time range). The second module, Session Builder, aggregates packets in flows (i.e., unidirectional flows identified by the classic 5-tuple), biflows (i.e.,
both directions of the traffic) or host sessions (aggregation of all the traffic of a
host). The F eature Extractor module calculates the features needed by the classification plugins. There is a single module for feature extraction in order to avoid
redundant calculations for different plugins. TIE provides a multi-classifier engine
divided in a Decision Combiner module and a set of classification plugins. On the
128
CHAPTER 7. OTHER IMPROVEMENTS TO TRAFFIC CLASSIFIERS
one hand, the Decision Combiner is in charge of calling several classification plugins when their features are available. On the other hand, this module merges the
results obtained from the different classification plugins in a definitive classification
result. In order to allow comparisons between different methods, the Output module provides the classification results from the Classif ication Combiner based on
a set of applications and groups of applications defined by the user.
TIE supports three different operating modes. The of f line mode generates
the classification results at the end of the TIE execution. The real-time mode
outputs the classification results as soon as possible, while the cycling mode is an
hybrid mode that generates the information every n minutes.
KD-Tree Plugin
In order to evaluate our traffic classification method, while providing a ready-touse tool for network operators, we implement the K-dimensional tree technique
as a TIE plugin. Before describing the details of this new plugin, we introduce
the K-dimensional tree technique. In particular, we focus on the major differences
with the original NN search algorithm.
The K-dimensional tree is a data structure to efficiently implement the Nearest Neighbor search algorithm. It represents a set of N points in K-dimensional
spaces as described by Friedman et al. [112] and Bentley [113]. These spaces are
maintained in a binary tree in which each non-leaf node generates a hyperplane
that divides the spaces into two subspaces.
The original NN algorithm searches iteratively the nearest point i, from a set
of points E, to a point p. In order to find the i point, it computes, for each point in
E, the distance (e.g., Euclidean or Manhattan distance) to the point p. Likewise,
if we are performing a K-NN search, the algorithm looks for the K i points nearest
to the point p. This search has O(N) time complexity and becomes unpractical
with the amount of traffic found in current networks.
On the contrary, the search in a K-dimensional tree allows to find in average
the nearest point in O(log N), with the additional cost of spending once O(N log
N) building the binary tree. Besides this notable improvement, the structure also
supports approximate searches, which can substantially improve the classification
time at the cost of producing a very small error.
The K-dimensional tree plugin that we implement in TIE is a combination of
the K-dimensional tree implemented in the C++ ANN library and a structure to
represent the relevant information still provided by the port numbers. In particular, we create an independent K-dimensional tree for each relevant port. We
7.2. EARLY CLASSIFICATION OF NETWORK TRAFFIC
129
refer as relevant ports as those that generate more traffic. Although the list of
relevant ports can be computed automatically, we also provide the user with the
option of manually configuring this list. Another configuration parameter is the
approximation value, which allows the method to improve its classification speed
by performing an approximate NN search. In the evaluation, we set this parameter to 0, which means that this approximation feature is not used. However,
higher values of this parameter could substantially improve the classification time
in critical scenarios, while still obtaining a reasonable accuracy.
Unlike in the original NN algorithm, the proposed method requires a lightweight
training phase to build the K-dimensional tree structure. Before building the data
structure, a sanitization process is performed on the training data. This procedure
removes the instances labeled as unknown from the training dataset assuming
that they have similar characteristics to other known flows. This assumption is
similar to that of ML clustering methods, where unlabeled instances are classified
according to their proximity in the feature space to those that are known. The
sanitization process also removes repeated or indistinguishable instances.
The traffic features used by our plugin are the destination port number and
the length of first n packets of a flow (without considering the TCP handshake).
By using only the first n packets, the plugin can classify the flows very quickly,
providing the network operator with the possibility of quickly reacting to the
classification results. In order to accurately classify short flows, the training phase
also accepts flows with less than n packets by filling the empty positions with null
coordinates.
Continuous Training System
In this section, we show the interaction of our KD-T ree plugin with the rest of the
TIE architecture, and describe the modifications done in TIE to allow our plugin
to continuously retrain itself.
Figure 7.6 shows the data flow of our continuous training system based on
TIE. The first three modules are used without any modification as found in the
original version of TIE. Besides the implementation of the new KD-T ree plugin,
we significantly modified the Decision Combiner module and the L7 plugin.
Our continuous training system follows the original TIE operation mode most
part of the time. Every packet is aggregated in bidirectional flows while its features
are calculated. When the traffic features needed by our plugin are available (i.e.,
n first packet sizes or the flow expires), the flow is classified by the KD-T ree
plugin. Although the method was tested with bidirectional flows, the current
130
CHAPTER 7. OTHER IMPROVEMENTS TO TRAFFIC CLASSIFIERS
implementation also supports the classification of unidirectional flows.
In order to automatically retrain our plugin, as the training data becomes
obsolete, we need a technique to set the ground-truth. TIE already provides the
L7 plugin, which implements a DPI technique originally used by TIE for validation
purposes. We modified the implementation of this plugin to continuously produce
training data (which includes flow labels - that is, the ground-truth - obtained
by a modified version of L7-filter [8]) for future trainings. While every flow is
sent to the KD-T ree plugin through the main path, the Decision Combiner
module applies flow sampling to the traffic, which is sent through a secondary
path to the L7 plugin. This secondary path is used to (i) set the ground-truth for
the continuous training system, (ii) continuously check the accuracy of the KDT ree plugin by comparing its output with that of L7, and (iii) keep the required
computational power low by using flow sampling (performing DPI on every single
flow will significantly decrease the performance of TIE).
The Decision Combiner module is also in charge of automatically triggering
the training of the KD-T ree plugin according to three different events that can
be configured by the user: after p packets, after s seconds, or if the accuracy of
the plugin compared to the L7 output is below a certain threshold t. The flows
classified by the L7 plugin, together with their features (i.e., destination port, n
packet sizes, L7 label), are placed in a queue. This queue keeps the last f classified
flows or the flows classified during the last s seconds.
The training module of the KD-T ree plugin is executed in a separate thread.
This way, the KD-T ree plugin can continuously classify the incoming flows without interruption, while it is periodically updated. In addition, it is possible to
automatically update the list of relevant ports by using the training data as a
reference.
7.2.3
Results
This section presents the performance evaluation of the proposed technique. First,
Subsection 7.2.3 describes the dataset used for the experiment. Subsection 7.2.3
performs a comparison between the original Nearest Neighbor algorithm and the
K-dimensional tree implementation. Subsection 7.2.3 presents a performance evaluation of the proposed plugin described in Subsection 7.2.2 evaluating different
aspects of the technique as the relevant ports or the number of packet sizes used
for the classification. Finally, Subsection 7.2.3 presents a preliminary study of the
impact of the continuous training system in the traffic classification.
7.2. EARLY CLASSIFICATION OF NETWORK TRAFFIC
131
Figure 7.6: Diagram of the Continuous Training Traffic Classification system based
on TIE
Evaluation Datasets
The evaluation dataset used in our performance evaluation consists of 8 fullpayload traces collected at the Gigabit access link of the Universitat Politècnica
de Catalunya BarcelonaTech. This dataset is derived from the UPC dataset described in Section 2.3.1. However, new traces are added and different ground-truth
labeling is applied.
Table 7.11 presents the details of the traces used in the evaluation. In order
to evaluate the proposed method, we used the first seven traces. Among those
traces we selected a single trace (UPC-II) as training dataset, which is the 15
min. long trace that contains the highest diversity in terms of instances from
different applications. We limit our training set to one trace in order to leave
a meaningful number of traces for the evaluation that are not used to build the
classification model. Therefore, the remaining first seven traces have been used as
the validation dataset. The last trace, UPC-VIII, was recorded with a difference in
time of four months with the trace UPC-II. Given this time difference we perform
a preliminary experiment with both traces in order to evaluate the gain provided
by our continuous training solution.
Nearest Neighbor vs K-dimensional Tree
In Section 7.2.2, we already discussed the main advantages of the K-dimensional
tree technique versus the original Nearest Neighbor algorithm. In order to present
numerical results showing this gain, we perform a comparison between both meth-
132
CHAPTER 7. OTHER IMPROVEMENTS TO TRAFFIC CLASSIFIERS
Table 7.11: Characteristics of the traffic traces in our dataset
Name
Date
Day
UPC-I
UPC-II
UPC-III
UPC-IV
UPC-V
UPC-VI
UPC-VII
UPC-VIII
11-12-08
11-12-08
12-12-08
12-12-08
14-12-08
21-12-08
22-12-08
10-03-09
Thu
Thu
Fri
Fri
Sun
Sun
Mon
Tue
Start
Time
10:00
12:00
01:00
16:00
00:00
12:00
12:30
03:00
Duration
15 min
15 min
15 min
15 min
15 min
1h
1h
1h
Valid
Flows
1 936 K
2 047 K
1 419 K
2 176 K
1 346 K
3 793 K
6 684 K
3 711 K
Bytes
53 G
63 G
38 G
55 G
29 G
133 G
256 G
78 G
Avg.
Util
482 Mbps
573 Mbps
345 Mbps
500 Mbps
263 Mbps
302 Mbps
582 Mbps
177 Mbps
ods. We evaluate our method with the original NN search implemented for validation purposes by the ANN library. Given that the ANN library implements both
methods in the same structure, we could not be able to show original memory
resources for the naive NN technique. We tested both methods with the trace
UPC-II (i.e. 500 000 flows after the sanitization process) using a 3GHz machine
with 4GB of RAM. It is important to note that given we are performing an offline evaluation we do not approximate the NN search neither in the NN original
algorithm or in the K-dimensional tree technique. For this reason the accuracy in
both methods is the same.
Table 7.12 summarizes the improvements obtained with the combination of the
K-dimensional tree technique with the port information. It presents the results in
terms of classifications per second depending on the number of packets sizes needed
for the classification and the list of relevant ports. There are three possible lists
of relevant ports. The unique list, where there is no relevant port and all the
instances belong to the same K-dimensional tree or NN structure. The selected
list, which is composed by the set of ports that contains most of the traffic from the
trace UPC-II. And the list where all the ports from the trace UPC-II are relevant.
The first column represents the original NN presented in previous works [17,34,36]
where all the information is represented in a single structure. When only one
packet is required, the proposed method is ten times faster than the original NN,
however the speed of the original method dramatically decreases when the number
of packets required increases becoming even a hundred times slower than the Kdimensional tree technique. In almost all the situations, the introduction of the
7.2. EARLY CLASSIFICATION OF NETWORK TRAFFIC
133
Table 7.12: Speed Comparison (flows/s) : Nearest Neighbor vs K-Dimensional
Tree
Packet
Size
1
5
7
10
Naive Nearest Neighbor
Unique Selected Ports All Ports
45 578
104 167
185 874
540
2 392
4 333
194
1 007
1 450
111
538
796
Unique
423 729
58 617
22 095
1 928
K-Dimensional Tree
Selected Ports All Ports
328 947
276 243
77 280
159 744
34 674
122 249
4 698
48 828
Table 7.13: Memory Comparison: Nearest Neighbor vs K-Dimensional Tree
Packet
Size
1
5
7
10
Naive
Nearest Neighbor
Unknown
Unknown
Unknown
Unknown
K-Dimensional Tree
Unique Selected Ports All Ports
40.65 MB
40.69 MB
40.72 MB
52.44 MB
52.63 MB
53.04 MB
56.00 MB
56.22 MB
57.39 MB
68.29 MB
68.56 MB
70.50 MB
relevant ports substantially increases the classification speed in both methods.
Tables 7.13 and 7.14 show the extremely low price that the K-dimensional tree
technique pays for a notable improvement in classification speed. The results show
that the memory resources required for our method are few. The memory used in
the K-dimensional tree is almost independent from the relevant ports parameter
and barely affected by the number of packet sizes. Regarding time, we believe that
the trade-off of the training phase is well compensated by the ability to use our
method as an online classifier. In the worst case, our method only takes about 20
seconds for the building phase.
Considering that both methods output the same classification results, the data
presented in this subsection show that the combination of the relevant ports and
the K-dimensional tree technique significantly improves the original NN search
with the only drawback of a (very fast) training phase but allowing us to use our
method as an efficient online traffic classifier.
134
CHAPTER 7. OTHER IMPROVEMENTS TO TRAFFIC CLASSIFIERS
Table 7.14: Building Time Comparative: Nearest Neighbor vs K-Dimensional Tree
Packet
Size
1
5
7
10
Naive
Nearest Neighbor
0s
0s
0s
0s
K-Dimensional Tree
Unique Selected Ports All Ports
13.01 s
12.72 s
12.52 s
16.45 s
16.73 s
15.62 s
17.34 s
16.74 s
16.07 s
19.81 s
19.59 s
18.82 s
K-dimensional tree plugin evaluation
In this section, we study the accuracy of our method depending on the different
parameters of the KD − T ree plugin. Figure 7.7a presents the accuracy of our
method by the number of packet sizes for the different traces of the dataset. In
this case, no information from the relevant ports is taken into account producing
a single K-dimensional tree. With this variation, using only the first two packets
we achieve an accuracy of almost 90%. The accuracy increases with the number
of packet sizes until a stable accuracy > 95% is reached with seven packet sizes.
In order to show the impact of using the list of relevant ports in the classification, in Figure 7.7b, we show the distribution of the first packet size for the training
trace UPC-II. Although there are some portions of the distribution dominated by a
group of applications, most of the applications have their first packet size between
the 0 and the 300 bytes ticks. This collision explains the poor accuracy presented
in the previous figure with only one packet.
The second parameter of our method, the relevant ports, furthermore of improve the classification speed appears to alleviate that situation. Figure 7.8a
present the accuracy of our method by number of packets using the set of relevant
ports that contains most of the traffic in UPC-II. With the help of the relevant
ports, our method achieves an accuracy > 90% using only the first packet size and
achieving a stable accuracy of 97% with seven packets.
Figure 7.8b presents the accuracy of our method depending on the set of
relevant ports with seven packet sizes. We choose seven because as can be seen in
the Figures 7.7a and 7.8a increasing the number of packet sizes for values greater
than seven do not improve its accuracy but decrease its classification speed. Using
all the ports of the training trace UPC-II, the method achieve the highest accuracy
with the own trace however classifying the rest of traces the accuracy substantially
7.2. EARLY CLASSIFICATION OF NETWORK TRAFFIC
135
100%
Accuracy
90%
80%
UPC−I
UPC−II
UPC−III
UPC−IV
UPC−V
UPC−VI
UPC−VII
70%
60%
50%
1
2
3
4
5
6
7
8
9
10
Number of Packet Sizes
(a) K-dimensional tree accuracy (by flow) without relevant ports
support, by number of packet sizes
100000
# Flows
10000
#
−
INTERACTIVE
GAME
P2P
FILE−SYSTEM
ENCRYPTED
TUNNELING
#
1000
−−
#−
−−
−−
−
−−
−
−−
−
−−
−
#− −
−# −
−
# −
− −
##− −
−−
#
−
−
# −
−
−
#
−−
−
− −
−#
#−
− #−−
−
−− −
− −− −
#−
−# −
#− −
− −−−−
##−
#−−#−
## #−−−−− −
#−
100
10
5
3
1
−300
WEB
MAIL
BULK
CONFERENCE
MULTIMEDIA
SERVICES
−
−
−−−−−−
−
−−#−
−# −
#−
#−#−
#−###−
#−#−
##
##
#−
#−−−−−−# −#−−−− − −#− − #−−− #
0
#
−
− −
300
##−− # # −− −
600
#
−
−
900
−
1200
−
1500
Packet Size
(b) First packet size distribution in the training trace UPC-II
Figure 7.7: K-dimensional tree evaluation without the support of the relevant
ports
decrease but being always higher than 85%. This decrease is given because using
all the ports as relevant ports is a very dependent option of the scenario and
could present classification inaccuracies with new instances belonging to ports not
represented in the training data. However, the figure shows that using a set of
relevant ports - in our case the ports that receive more than 1% of the traffic besides increasing the classification speed also improves accuracy.
Erman et al. in [27] pointed out a common situation found among the ML
techniques: the accuracy when measured by flows is much higher than when measured by bytes or packets. This usually happens because some elephant-flows are
not correctly classified. Figures 7.9a and 7.9b present the classification results
136
CHAPTER 7. OTHER IMPROVEMENTS TO TRAFFIC CLASSIFIERS
100%
Accuracy
90%
80%
UPC−I
UPC−II
UPC−III
UPC−IV
UPC−V
UPC−VI
UPC−VII
70%
60%
50%
1
2
3
4
5
6
7
8
9
10
Number of Packet Sizes
(a) K-dimensional tree accuracy (by flow) with relevant ports support, by number of packet sizes
Accuracy
100%
95%
90%
All
Single
Selected
85%
UPC−I
UPC−II
UPC−III
UPC−IV
UPC−V
UPC−VI
UPC−VII
Traces
(b) K-dimensional tree accuracy (by flow) by set of relevant ports
with a fixed number of packet sizes(i,e,. 7)
Figure 7.8: K-dimensional tree evaluation with the support of the relevant ports
of our method considering also the accuracy by bytes and packets, showing that,
unlike other ML solutions, our method is able to keep high accuracy values even
with such metrics. This is because our method is very accurate with the group
of applications P2P and Web, the groups that represent in terms of bytes most of
the traffic in our traces.
Finally we also study the accuracy of our method regarding the groups of applications. In our evaluation, we use the original definition of the applications and
group of applications given in TIE. Figure 7.10 shows that our method is able to
classify with excellent accuracy the most popular groups of applications. However,
the accuracy of the applications groups that are not very common substantially
7.2. EARLY CLASSIFICATION OF NETWORK TRAFFIC
137
Accuracy
100%
95%
90%
All
Single
Selected
85%
UPC−I
UPC−II
UPC−III
UPC−IV
UPC−V
UPC−VI
UPC−VII
Traces
(a) K-dimensional tree accuracy (by packet) by set of relevant ports
with a fixed number of packet sizes(i,e,. 7)
Accuracy
100%
95%
90%
All
Single
Selected
85%
UPC−I
UPC−II
UPC−III
UPC−IV
UPC−V
UPC−VI
UPC−VII
Traces
(b) K-dimensional tree accuracy (by byte) by set of relevant ports
with a fixed number of packet sizes(i,e,. 7)
Figure 7.9: K-dimensional tree evaluation with the support of the relevant ports
decreases. These accuracies have a very low impact on the final accuracy of the
method given that the representation of these groups in the traces used is almost
negligible. A possible solution to improve the accuracy for these groups of applications could be the addition of artificial instances of these groups in the training
data.
Definitely, we present a set of results showing how the K-dimensional tree technique using the still useful information provided by the ports improves almost all
the aspects of previous methods based in the NN search. With the unique drawback of a short training phase, our method is able to perform online classification
with a very high accuracy, > 90% with only one packet or > 97% with seven
138
CHAPTER 7. OTHER IMPROVEMENTS TO TRAFFIC CLASSIFIERS
100%
UPC−I
UPC−III
UPC−IV
UPC−V
UPC−VI
UPC−VII
AVERAGE
90%
80%
Accuracy
70%
60%
50%
40%
30%
20%
10%
0%
CONFERENCING
P2P
WEB
SERVICES
ENCRYPTION
GAMES
MAIL
MULTIMEDIA
BULK
FILE_SYSTEM
TUNNEL
INTERACTIVE
Application Groups
Figure 7.10: Accuracy by application group (seven packet sizes and selected list
of ports as parameters)
packets.
Continuous Training System Evaluation
This section presents a preliminary study of the impact of our continuous training
traffic classifier. Due to lack of traces comprising a very long period of time and
because of the intrinsic difficulties in processing such large traces, we simulate a
scenario in which the features of the traffic evolve by concatenating the UPC-II
and UPC-VIII traces. The trace UPC-VIII, besides belonging to a difference daytime, was recorded four months later than UPC-II, this suggests a different traffic
mix with different properties. Using seven as the fixed number of packets sizes,
the results in Table 7.15 confirm our intuition. On the one hand, using the trace
UPC-II as training data to classify the trace UPC-VIII, we obtain an accuracy of
almost 85%. On the other hand, after detecting such decrease of accuracy and
thus performing retraining using the first fifteen minutes of the trace UPC-VIII,
we obtain and impressive accuracy of 98,17% showing the important impact of the
continuous training.
The results of a second experiment are also presented in Table 7.15. Instead of
retraining the system with a new training data, we study if the modification of the
list of relevant ports is enough to obtain the original accuracy. The results show
that this solution does not bring any improvement when applied alone. However,
7.2. EARLY CLASSIFICATION OF NETWORK TRAFFIC
139
Table 7.15: Evaluation of the Continuous Training system by training trace and
set of relevant ports
Training Trace
Relevant Port List
Accuracy
UPC-II
UPC-II UPC-VIII
84.20 %
76.10 %
First 15 min. UPC-VIII
UPC-II
UPC-VIII
98.17 %
98.33 %
the optimum solution is obtained when both the training data and the list of
relevant ports are updated and the system is then retrained.
7.2.4
Related Work
The NN method for traffic classification was firstly proposed in [17], where a
comparison of the NN technique with the Linear Discriminant Analysis method
was presented. They showed that NN was able to classify, among seven different
classes of traffic, with an error rate below 10%.
However, the most interesting conclusions about the NN algorithm are found
in the works from Williams et al. [34] and Kim et al. [36]. Both works compared
different ML methods and showed the pros and cons of the NN algorithm for traffic classification. In summary, NN was shown to be one of the most accurate ML
methods, with the additional feature of requiring zero time to build the classification model. However, NN was the ML-based algorithm with the worst results
in terms of classification speed. This is the reason why NN is often discarded for
online classification.
The efficient implementation of the NN algorithm presented in this paper is
based instead on the K-dimensional tree, which solves its problems in terms of
classification speed, while keeping very high accuracy. Another important feature
of our method is its ability to early classify the network traffic. This idea is
exported from the work from Bernaille et al. [20, 21]. This early classification
feature allows the method to classify the network traffic by just using the first
packets of each flow. Bernaille et al. compared three different unsupervised ML
methods (K-Means, GMM and HMM), while in this work we apply this idea to a
supervised ML method (NN).
140
CHAPTER 7. OTHER IMPROVEMENTS TO TRAFFIC CLASSIFIERS
7.2.5
Section Summary
In this work, we revisited the viability of using the Nearest Neighbor algorithm
(NN) for online traffic classification, which has been often discarded in previous
studies due to its poor classification speed. In order to address this well-known
limitation, we presented an efficient implementation of the NN algorithm based
on a K-dimensional tree data structure, which can be used for online traffic classification with high accuracy and low overhead. In addition, we combined this
technique with the relevant information still provided by the port numbers, which
further increases its classification speed and accuracy.
Our results show that our method can achieve very high accuracy (> 90%) by
looking only at the first packet of a flow. When the number of analyzed packets
is increased to seven, the accuracy of our method increases beyond 95%. This
early classification feature is very important, since it allows network operators to
quickly react to the classification results.
We presented an actual implementation of our traffic classification method
based on the TIE classification engine. The main novelty of our implementation
is its continuous training feature, which allows the system to be automatically retrained by itself as the training data becomes obsolete. Our preliminary evaluation
of this unique feature presents very encouraging results.
Chapter 8
Conclusions
This thesis has addressed some of the most important challenges in the deployment and maintenance of traffic classification solutions in production networks.
We focused on three main aspects of network traffic classifiers that hinder the
introduction of existing techniques in production networks: (i) the ease of deployment, (ii) the ease of maintenance and (iii) the validation and comparison
of existing techniques. Furthermore, to reduce the lack of adequate data to the
research community, we published several datasets used in our evaluation to allow
the fair validation and comparison of current and future classification techniques.
First, to facilitate the deployment of existing techniques in production networks
we:
1. addressed the traffic classification problem with NetFlow data using a wellknown supervised learning technique. Our results show that supervised
methods (e.g., C4.5) can achieve high accuracy (≈90%) with unsampled
NetFlow data, despite the limited information provided by NetFlow, as compared to the packet-level data used in previous studies.
2. empirically and analytically analyzed the severe impact of packet sampling
on the classification accuracy of supervised learning methods.
3. proposed a simple improvement in the training process that significantly
increased the accuracy of our classification method under packet sampling.
For example, for a sampling rate of 1/100, we achieved an overall accuracy of
85% in terms of classified flows, while before we could not reach an accuracy
beyond 50%.
141
142
CHAPTER 8. CONCLUSIONS
Given that Sampled NetFlow is the most widely extended monitoring solution
among network operators, we sincerely think that based on our results, using
Sampled NetFlow as input considerably facilitates the deployment of existing techniques in operational networks.
Second, aiming to the address the maintenance problem of existing techniques
in operational networks we:
1. showed that classification models suffer from temporal and spatial obsolescence.
2. addressed this problem by first proposing a complete traffic classification
solution with a novel automatic retraining system. The classification system
combines the advantages of three different techniques (i.e., IP-based, Servicebased and ML-based) along with the autonomic retraining system to sustain
a high classification accuracy during long periods of time. The retraining
system combines multiple DPI techniques and only requires a small sample
of the whole traffic to keep the classification system updated.
3. introduced the use of stream-based ML techniques for network traffic classification by proposing a new stream-based classification solution based on
Hoeffding Adaptive Trees. The main novelty of our technique is its ability
to automatically adapt to the changes of the traffic with just a small sample
of labeled data, making our solution very easy to maintain. This solution is
not only more accurate than traditional batch-based techniques, but it also
sustains high accuracy over the years with less cost.
Our proposals have several features that are particularly appealing for network
management: (i) high classification accuracy and completeness, (ii) support for
NetFlow data, (iii) automatic model retraining, and (iv) resilience to sampling.
These features altogether result in a significant reduction in the cost of deployment,
operation and maintenance compared to previous methods based on packet traces
and manually-made classification models.
Third, to present a first step towards the fair validation and comparison of
existing network traffic classifiers we:
1. created a reliable labeled dataset with full packet payload. The dataset of
more than 500 K flows contains traffic from popular applications like HTTP,
eDonkey, BitTorrent, FTP, DNS, NTP, RDP, NETBIOS, SSH, and RDP.
The total amount of data properly labeled is 32.61 GB. We released this
dataset to the research community with full payload, so it can be used as
8.1. FUTURE WORK
143
a common reference for the comparison and validation of network traffic
classifiers.
2. using the previous dataset, we compared the performance of six tools (i.e.,
PACE, OpenDPI, L7-filter, nDPI, Libprotoident, and NBAR), which are
usually used for ground-truth generation. The results obtained show that
PACE is, on our dataset, the most reliable solution for ground-truth generation. Among the open-source tools, nDPI and especially Libprotoident
present the best results. On the other hand, NBAR and L7-filter present
several inaccuracies that make them not recommendable as a ground-truth
generator.
All in all, we truly believe the different contributions obtained in this thesis
will help to reduce the gap between the real-world requirements from the network
industry, and the research being carried out in the field of network traffic classification. The results presented in this thesis give useful indications to researchers
and developers to design future accurate and realistic network traffic classification
solutions for production networks.
8.1
Future Work
In this thesis, we focused on addressing practical challenges of network traffic
classification solutions. Moreover, due the continuous evolution of the Internet
traffic and its applications the problems we tackled allow further research and
improvements.
To maintain the classification solution updated, we relied on the labeling of a
small sample of the traffic by DPI-based techniques. However, DPI-based techniques should be also periodically updated. Although, this is usually provided
by DPI tools developers, some techniques can help us to detect when an update
of the ground-truth generator is necessary. For instance, the detection of new
applications not classified by the ground-truth generator could be obtained using
unsupervised machine learning techniques.
Another future work derived from the continuous evolution of the Internet
traffic is the extension of the publicly available reliable labeled datasets. In order to allow for future comparison and validation of new proposals, new datasets
containing new applications traffic should be provided to the research community.
There is also a wide range of applications following the early classification
approach. For example, we can use this early classification to apply traffic shaping
144
CHAPTER 8. CONCLUSIONS
based on the identified application. However, the application of traffic shaping is
a very controversial topic given its implications in network neutrality.
Finally, an interesting approach barely studied is the multilabel classification.
Most of the existing solutions for traffic classification focus on the classification of
the traffic at a single level. However, network traffic can be classified at different
levels. For example, the traffic produced by a person watching a Youtube video
in his cell phone can be considered correctly classified as: Youtube traffic, RTSP
traffic, Streaming traffic, Flash traffic, UDP traffic or Mobile traffic. It would
be interesting to classify the traffic at different levels trying to be as complete as
possible.
Bibliography
[1] T. Karagiannis, A. Broido, M. Faloutsos, and K. claffy, “Transport
layer identification of p2p traffic,” in Proceedings of the 4th ACM
SIGCOMM Conference on Internet Measurement, ser. IMC ’04. New
York, NY, USA: ACM, 2004, pp. 121–134. [Online]. Available: http:
//doi.acm.org/10.1145/1028788.1028804
[2] Service Name and Transport Protocol Port Number Registry (IANA), [Online]. Available: http://www.iana.org/assignments/port-numbers, as of October 17, 2014.
[3] R. Beverly, S. Bauer, and A. Berger, “The internet is not a big
truck: Toward quantifying network neutrality,” in Proceedings of the 8th
International Conference on Passive and Active Network Measurement, ser.
PAM’07. Berlin, Heidelberg: Springer-Verlag, 2007, pp. 135–144. [Online].
Available: http://dl.acm.org/citation.cfm?id=1762888.1762907
[4] T. Karagiannis, A. Broido, N. Brownlee, K. Claffy, and M. Faloutsos, “Filesharing in the internet: a characterization of p2p traffic in the backbone,”
Univ. of California, Riverside, USA, Technical Report, 2003.
[5] ——, “Is p2p dying or just hiding?” in Global Telecommunications Conference. IEEE GLOBECOM’04, vol. 3, 2004, pp. 1532–1538.
[6] A. W. Moore and K. Papagiannaki, “Toward the accurate identification of
network applications,” in Proceedings of the 6th International Conference
on Passive and Active Network Measurement, ser. PAM’05. Berlin,
Heidelberg:
Springer-Verlag, 2005, pp. 41–54. [Online]. Available:
http://dx.doi.org/10.1007/978-3-540-31966-5 4
[7] S. Sen, O. Spatscheck, and D. Wang, “Accurate, scalable in-network
identification of p2p traffic using application signatures,” in Proceedings
145
146
BIBLIOGRAPHY
of the 13th International Conference on World Wide Web, ser. WWW
’04. New York, NY, USA: ACM, 2004, pp. 512–521. [Online]. Available:
http://doi.acm.org/10.1145/988672.988742
[8] L7-filter: Application layer packet classifier for Linux., [Online]. Available:
http://l7-filter.sourceforge.net/, as of October 31, 2014.
[9] OpenDPI: the Open Source version of ipoque’s DPI engine, [Online]. Available: http://www.opendpi.org/, as of October 31, 2014.
[10] PACE, ipoque‘s Protocol and Application Classification Engine, [Online].
Available: http://www.ipoque.com/en/products/pace, as of October 31,
2014.
[11] nDPI, Open and Extensible LGPLv3 Deep Packet Inspection Library, [Online]. Available: http://www.ntop.org/products/ndpi/, as of October 31,
2014.
[12] T. Auld, A. W. Moore, and S. F. Gull, “Bayesian neural networks for internet
traffic classification,” Trans. Neur. Netw., vol. 18, no. 1, pp. 223–239, Jan.
2007. [Online]. Available: http://dx.doi.org/10.1109/TNN.2006.883010
[13] M. Crotti, M. Dusi, F. Gringoli, and L. Salgarelli, “Traffic classification
through simple statistical fingerprinting,” SIGCOMM Comput. Commun.
Rev., vol. 37, no. 1, pp. 5–16, Jan. 2007. [Online]. Available:
http://doi.acm.org/10.1145/1198255.1198257
[14] P. Haffner, S. Sen, O. Spatscheck, and D. Wang, “Acas: Automated
construction of application signatures,” in Proceedings of the 2005 ACM
SIGCOMM Workshop on Mining Network Data, ser. MineNet ’05.
New York, NY, USA: ACM, 2005, pp. 197–202. [Online]. Available:
http://doi.acm.org/10.1145/1080173.1080183
[15] H. Jiang, A. W. Moore, Z. Ge, S. Jin, and J. Wang, “Lightweight
application classification for network management,” in Proceedings of the
2007 SIGCOMM Workshop on Internet Network Management, ser. INM
’07. New York, NY, USA: ACM, 2007, pp. 299–304. [Online]. Available:
http://doi.acm.org/10.1145/1321753.1321771
[16] A. W. Moore and D. Zuev, “Internet traffic classification using bayesian
analysis techniques,” in Proceedings of the 2005 ACM SIGMETRICS
BIBLIOGRAPHY
147
International Conference on Measurement and Modeling of Computer
Systems, ser. SIGMETRICS ’05. New York, NY, USA: ACM, 2005, pp.
50–60. [Online]. Available: http://doi.acm.org/10.1145/1064212.1064220
[17] M. Roughan, S. Sen, O. Spatscheck, and N. Duffield, “Class-of-service
mapping for qos: A statistical signature-based approach to ip traffic
classification,” in Proceedings of the 4th ACM SIGCOMM Conference on
Internet Measurement, ser. IMC ’04. New York, NY, USA: ACM, 2004, pp.
135–148. [Online]. Available: http://doi.acm.org/10.1145/1028788.1028805
[18] D. Zuev and A. W. Moore, “Traffic classification using a statistical
approach,” in Proceedings of the 6th International Conference on
Passive and Active Network Measurement, ser. PAM’05. Berlin,
Heidelberg: Springer-Verlag, 2005, pp. 321–324. [Online]. Available:
http://dx.doi.org/10.1007/978-3-540-31966-5 25
[19] G. Szabó, I. Szabó, and D. Orincsay, “Accurate traffic classification,” in
World of Wireless, Mobile and Multimedia Networks. WoWMoM 2007. IEEE
International Symposium on a, 2007, pp. 1–8.
[20] L. Bernaille, R. Teixeira, and K. Salamatian, “Early application
identification,” in Proceedings of the 2006 ACM CoNEXT Conference, ser.
CoNEXT ’06. New York, NY, USA: ACM, 2006, pp. 6:1–6:12. [Online].
Available: http://doi.acm.org/10.1145/1368436.1368445
[21] L. Bernaille, R. Teixeira, I. Akodkenou, A. Soule, and K. Salamatian,
“Traffic classification on the fly,” SIGCOMM Comput. Commun.
Rev., vol. 36, no. 2, pp. 23–26, Apr. 2006. [Online]. Available:
http://doi.acm.org/10.1145/1129582.1129589
[22] L. Bernaille and R. Teixeira, “Early recognition of encrypted applications,”
in Proceedings of the 8th International Conference on Passive and Active
Network Measurement, ser. PAM’07. Berlin, Heidelberg: Springer-Verlag,
2007, pp. 165–175. [Online]. Available: http://dl.acm.org/citation.cfm?id=
1762888.1762911
[23] A. Dainotti, W. De Donato, A. Pescape, and P. Salvo Rossi, “Classification of network traffic via packet-level hidden markov models,” in Global
Telecommunications Conference. IEEE GLOBECOM’08, 2008, pp. 1–5.
148
BIBLIOGRAPHY
[24] J. Erman, M. Arlitt, and A. Mahanti, “Traffic classification using clustering
algorithms,” in Proceedings of the 2006 SIGCOMM Workshop on Mining
Network Data, ser. MineNet ’06. New York, NY, USA: ACM, 2006, pp.
281–286. [Online]. Available: http://doi.acm.org/10.1145/1162678.1162679
[25] J. Erman, A. Mahanti, M. Arlitt, I. Cohen, and C. Williamson,
“Offline/realtime traffic classification using semi-supervised learning,”
Perform. Eval., vol. 64, no. 9-12, pp. 1194–1213, Oct. 2007. [Online].
Available: http://dx.doi.org/10.1016/j.peva.2007.06.014
[26] J. Erman, A. Mahanti, M. Arlitt, and C. Williamson, “Identifying and
discriminating between web and peer-to-peer traffic in the network core,”
in Proceedings of the 16th International Conference on World Wide Web,
ser. WWW ’07. New York, NY, USA: ACM, 2007, pp. 883–892. [Online].
Available: http://doi.acm.org/10.1145/1242572.1242692
[27] J. Erman, A. Mahanti, and M. Arlitt, “Byte me: A case for byte accuracy
in traffic classification,” in Proceedings of the 3rd Annual ACM Workshop on
Mining Network Data, ser. MineNet ’07. New York, NY, USA: ACM, 2007,
pp. 35–38. [Online]. Available: http://doi.acm.org/10.1145/1269880.1269890
[28] A. Soule, K. Salamatia, N. Taft, R. Emilion, and K. Papagiannaki, “Flow
classification by histograms: Or how to go on safari in the internet,”
in Proceedings of the Joint International Conference on Measurement
and Modeling of Computer Systems, ser. SIGMETRICS ’04/Performance
’04. New York, NY, USA: ACM, 2004, pp. 49–60. [Online]. Available:
http://doi.acm.org/10.1145/1005686.1005696
[29] S. Zander, T. Nguyen, and G. Armitage, “Self-learning ip traffic
classification based on statistical flow characteristics,” in Proceedings of the
6th International Conference on Passive and Active Network Measurement,
ser. PAM’05. Berlin, Heidelberg: Springer-Verlag, 2005, pp. 325–328.
[Online]. Available: http://dx.doi.org/10.1007/978-3-540-31966-5 26
[30] ——, “Automated traffic classification and application identification using
machine learning,” in Local Computer Networks. 30th Anniversary. The
IEEE Conference on, 2005, pp. 250–257.
[31] J. Erman, A. Mahanti, and M. Arlitt, “Internet traffic identification using machine learning,” in Global Telecommunications Conference. GLOBECOM’06. IEEE, 2006, pp. 1–6.
BIBLIOGRAPHY
149
[32] N. Williams, S. Zander, and G. Armitage, “A preliminary performance
comparison of five machine learning algorithms for practical ip traffic flow
classification,” SIGCOMM Comput. Commun. Rev., vol. 36, no. 5, pp. 5–16,
Oct. 2006. [Online]. Available: http://doi.acm.org/10.1145/1163593.1163596
[33] ——, “Evaluating machine learning methods for online game traffic identification,” CAIA Technical Report, April, 2006.
[34] ——, “Evaluating machine learning algorithms for automated network application identification,” CAIA Technical Report, April, 2006.
[35] G. Szabó, D. Orincsay, S. Malomsoky, and I. Szabó, “On the validation
of traffic classification algorithms,” in Proceedings of the 9th International
Conference on Passive and Active Network Measurement, ser. PAM’08.
Berlin, Heidelberg: Springer-Verlag, 2008, pp. 72–81. [Online]. Available:
http://dl.acm.org/citation.cfm?id=1791949.1791960
[36] Y.-s. Lim, H.-c. Kim, J. Jeong, C.-k. Kim, T. T. Kwon, and Y. Choi,
“Internet traffic classification demystified:
On the sources of the
discriminative power,” in Proceedings of the 6th International COnference,
ser. Co-NEXT ’10. New York, NY, USA: ACM, 2010, pp. 9:1–9:12.
[Online]. Available: http://doi.acm.org/10.1145/1921168.1921180
[37] T. T. Nguyen and G. Armitage, “A survey of techniques for
internet traffic classification using machine learning,” Commun. Surveys
Tuts., vol. 10, no. 4, pp. 56–76, Oct. 2008. [Online]. Available:
http://dx.doi.org/10.1109/SURV.2008.080406
[38] T. Karagiannis, K. Papagiannaki, and M. Faloutsos, “Blinc: Multilevel
traffic classification in the dark,” SIGCOMM Comput. Commun.
Rev., vol. 35, no. 4, pp. 229–240, Aug. 2005. [Online]. Available:
http://doi.acm.org/10.1145/1090191.1080119
[39] K. Xu, Z.-L. Zhang, and S. Bhattacharyya, “Profiling internet backbone
traffic: Behavior models and applications,” SIGCOMM Comput. Commun.
Rev., vol. 35, no. 4, pp. 169–180, Aug. 2005. [Online]. Available:
http://doi.acm.org/10.1145/1090191.1080112
[40] T. Karagiannis, K. Papagiannaki, N. Taft, and M. Faloutsos, “Profiling
the end host,” in Proceedings of the 8th International Conference
150
BIBLIOGRAPHY
on Passive and Active Network Measurement, ser. PAM’07. Berlin,
Heidelberg: Springer-Verlag, 2007, pp. 186–196. [Online]. Available:
http://dl.acm.org/citation.cfm?id=1762888.1762913
[41] T. Mori, R. Kawahara, H. Hasegawa, and S. Shimogawa, “Characterizing
traffic flows originating from large-scale video sharing services,” in
Proceedings of the Second International Conference on Traffic Monitoring
and Analysis, ser. TMA’10. Berlin, Heidelberg: Springer-Verlag, 2010, pp.
17–31. [Online]. Available: http://dx.doi.org/10.1007/978-3-642-12365-8 2
[42] S.-H. Yoon, J.-W. Park, J.-S. Park, Y.-S. Oh, and M.-S. Kim,
“Internet application traffic classification using fixed ip-port,” in
Proceedings of the 12th Asia-Pacific Network Operations and Management
Conference on Management Enabling the Future Internet for Changing
Business and New Computing Services, ser. APNOMS’09. Berlin,
Heidelberg:
Springer-Verlag, 2009, pp. 21–30. [Online]. Available:
http://dl.acm.org/citation.cfm?id=1814002.1814006
[43] QosMos ixEngine:
Deep Packet Inspection and Metadata Engine,
[Online].
Available:
http://www.qosmos.com/products/
deep-packet-inspection-engine/, as of October 31, 2014.
[44] Palo Alto Networks App-Id: Identifying any application on any port, [Online]. Available: https://www.paloaltonetworks.com/products/technologies/
app-id.html, as of October 31, 2014.
[45] Emulex, EndaceDAG Data Capture Cards, [Online]. Available: http:
//www.emulex.com/products/network-visibility-products-and-services/
endacedag-data-capture-cards, as of October 31, 2014.
[46] V. Carela-Español, P. Barlet-Ros, A. Cabellos-Aparicio, and J. Solé-Pareta,
“Analysis of the impact of sampling on netflow traffic classification,”
Comput. Netw., vol. 55, no. 5, pp. 1083–1099, Apr. 2011. [Online]. Available:
http://dx.doi.org/10.1016/j.comnet.2010.11.002
[47] W. Li, M. Canini, A. W. Moore, and R. Bolla, “Efficient application
identification and the temporal and spatial stability of classification
schema,” Comput. Netw., vol. 53, no. 6, pp. 790–809, Apr. 2009. [Online].
Available: http://dx.doi.org/10.1016/j.comnet.2008.11.016
BIBLIOGRAPHY
151
[48] S. Valenti, D. Rossi, A. Dainotti, A. Pescapè, A. Finamore, and
M. Mellia, “Datatraffic monitoring and analysis,” E. Biersack, C. Callegari,
and M. Matijasevic, Eds. Berlin, Heidelberg: Springer-Verlag, 2013,
ch. Reviewing Traffic Classification, pp. 123–147. [Online]. Available:
http://dl.acm.org/citation.cfm?id=2555672.2555680
[49] K. Fukuda, “Difficulties of identifying application type in backbone traffic,”
in Network and Service Management (CNSM), International Conference on.
IEEE, 2010, pp. 358–361.
[50] Alcock, S. and Nelson, R., “Libprotoident: Traffic Classification Using
Lightweight Packet Inspection,” University of Waikato, Tech. Rep., 2012,
[Online]. Available: http://www.wand.net.nz/publications/lpireport, as of
October 31, 2014.
[51] A. Dainotti, F. Gargiulo, L. I. Kuncheva, A. Pescape, and C. Sansone, “Identification of traffic flows hiding behind tcp port 80,” in Communications
(ICC), IEEE International Conference on, 2010, pp. 1–6.
[52] CoralReef, [Online]. Available: http://www.caida.org/tools/measurement/
coralreef/, as of October 31, 2014.
[53] G. Iannaccone, “Fast prototyping of network data mining applications,” in
Proceedings of the 7th International Conference on Passive and Active Network Measurement, ser. PAM’06, 2006.
[54] Cisco Systems: Sampled NetFlow., [Online]. Available: http://www.cisco.
com/c/en/us/td/docs/ios/12 0s/feature/guide/12s sanf.html, as of October
31, 2014.
[55] IPFIX: IP Flow Information Export., [Online]. Available: http://www.ietf.
org/html.charters/ipfix-charter.html, as of October 31, 2014.
[56] X. Tian, Q. Sun, X. Huang, and Y. Ma, “Dynamic online traffic
classification using data stream mining,” in Proceedings of the 2008
International Conference on MultiMedia and Information Technology, ser.
MMIT ’08. Washington, DC, USA: IEEE Computer Society, 2008, pp.
104–107. [Online]. Available: http://dx.doi.org/10.1109/MMIT.2008.185
[57] ——, “A dynamic online traffic classification methodology based on data
stream mining,” in Proceedings of the 2009 WRI World Congress on
152
BIBLIOGRAPHY
Computer Science and Information Engineering - Volume 01, ser. CSIE
’09. Washington, DC, USA: IEEE Computer Society, 2009, pp. 298–302.
[Online]. Available: http://dx.doi.org/10.1109/CSIE.2009.904
[58] B. Raahemi, W. Zhong, and J. Liu, “Peer-to-peer traffic identification by
mining ip layer data streams using concept-adapting very fast decision
tree,” in Proceedings of the 2008 20th IEEE International Conference on
Tools with Artificial Intelligence - Volume 01, ser. ICTAI ’08. Washington,
DC, USA: IEEE Computer Society, 2008, pp. 525–532. [Online]. Available:
http://dx.doi.org/10.1109/ICTAI.2008.12
[59] MAWI Working Group Traffic Archive, [Online]. Available: http://mawi.
wide.ad.jp/mawi/, as of October 31, 2014.
[60] Bujlow, T. et al., “Volunteer-Based System for classification of traffic in
computer networks,” in 19th Telecommunications Forum TELFOR. IEEE,
2011, pp. 210–213.
[61] V. Carela-Español, P. Barlet-Ros, O. Mula-Valls, and J. Sole-Pareta, “An
automatic traffic classification system for network operation and management,” Journal of Network and Systems Management, 2013.
[62] V. Carela-Español, T. Bujlow, and P. Barlet-Ros, “Is our ground-truth
for traffic classification reliable?” in Proceedings of the 15th International
Conference on Passive and Active Network Measurement, ser. PAM’14.
Springer, 2014, pp. 98–108.
[63] J. Molina Rodriguez, V. Carela-Español, P. Barlet Ros, R. Hoffmann, and
K. Degner, “Empirical analysis of traffic to establish a profiled flow termination timeout,” in Wireless Communications and Mobile Computing Conference (IWCMC), 9th International. IEEE, 2013, pp. 1156–1161.
[64] V. Carela-Español, P. Barlet-Ros, M. Solé-Simó, A. Dainotti, W. de Donato,
and A. Pescapé, “K-dimensional trees for continuous traffic classification,” in
Proceedings of the Second International Conference on Traffic Monitoring and
Analysis, ser. TMA’10. Berlin, Heidelberg: Springer-Verlag, 2010, pp. 141–
154. [Online]. Available: http://dx.doi.org/10.1007/978-3-642-12365-8 11
[65] D. Antoniades, M. Polychronakis, S. Antonatos, E. P. Markatos, S. Ubik, and
A. Øslebø, “Appmon: An application for accurate per application network
traffic characterization,” in In IST Broadband Europe 2006 Conference, 2006.
BIBLIOGRAPHY
153
[66] J. Cohen, “A coefficient of agreement for nominal scales,” Educ. and Psychol.
Meas., vol. 20, no. 1, pp. 37–46, 1960.
[67] M. Canini, W. Li, A. W. Moore, and R. Bolla, “Gtvs: Boosting the
collection of application traffic ground truth,” in Proceedings of the First
International Workshop on Traffic Monitoring and Analysis, ser. TMA ’09.
Berlin, Heidelberg: Springer-Verlag, 2009, pp. 54–63. [Online]. Available:
http://dx.doi.org/10.1007/978-3-642-01645-5 7
[68] Traffic classification at the Universitat Politècnica de Catalunya (UPC
BarcelonaTech), [Online]. Available: http://www.cba.upc.edu/monitoring/
traffic-classification, as of October 31, 2014.
[69] CAIDA Data - Overview of Datasets, Monitors, and Reports, [Online]. Available: http://www.caida.org/data/overview/, as of October 31, 2014.
[70] Waikato Internet Traffic Storage (WITS), [Online]. Available: http://wand.
net.nz/wits/, as of October 31, 2014.
[71] J. Quinlan, C4. 5: programs for machine learning.
1993.
Morgan Kaufmann,
[72] WEKA: data mining software in Java., [Online]. Available: http://www.cs.
waikato.ac.nz/ml/weka/, as of October 31, 2014.
[73] P. Barlet-Ros, J. Sole-Pareta, J. Barrantes, E. Codina, and J. DomingoPascual, “SMARTxAC: a passive monitoring and analysis system for highspeed networks,” Campus-Wide Information Systems, vol. 23, no. 4, 2006.
[74] K.-c. Lan and J. Heidemann, “A measurement study of correlations of
internet flow characteristics,” Comput. Netw., vol. 50, no. 1, pp. 46–62, Jan.
2006. [Online]. Available: http://dx.doi.org/10.1016/j.comnet.2005.02.008
[75] W. Liu, “Estimating flow length distributions using least square method and
maximum likelihood estimation,” in ICCS ’07: Proc. of the 7th international
conference on Computational Science, Part IV. Springer-Verlag, 2007, pp.
793–796.
[76] N. V. Chawla, “C4.5 and imbalanced data sets: investigating the effect of
sampling method, probabilistic estimate, and decision tree structure,” in
Proc. of the ICML’03 Workshop on Class Imbalances, 2003.
154
BIBLIOGRAPHY
[77] Cisco IOS NetFlow, [Online]. Available: http://www.cisco.com/c/en/us/
products/ios-nx-os-software/ios-netflow/index.html, as of October 31, 2014.
[78] H. Pham, Handbook of Reliability Engineering. Springer London, 2003.
[79] N. Duffield, C. Lund, and M. Thorup, “Properties and prediction of
flow statistics from sampled packet streams,” in Proceedings of the 2Nd
ACM SIGCOMM Workshop on Internet Measurment, ser. IMW ’02.
New York, NY, USA: ACM, 2002, pp. 159–171. [Online]. Available:
http://doi.acm.org/10.1145/637201.637225
[80] J. Mai, A. Sridharan, C.-N. Chuah, H. Zang, and T. Ye, “Impact
of packet sampling on portscan detection,” IEEE J.Sel. A. Commun.,
vol. 24, no. 12, pp. 2285–2298, Dec. 2006. [Online]. Available:
http://dx.doi.org/10.1109/JSAC.2006.884027
[81] J. Mai, C.-N. Chuah, A. Sridharan, T. Ye, and H. Zang, “Is sampled
data sufficient for anomaly detection?”
in Proceedings of the 6th
ACM SIGCOMM Conference on Internet Measurement, ser. IMC ’06.
New York, NY, USA: ACM, 2006, pp. 165–176. [Online]. Available:
http://doi.acm.org/10.1145/1177080.1177102
[82] I. Paredes-Oliva, P. Barlet-Ros, and J. Solé-Pareta, “Portscan detection
with sampled netflow,” in Proceedings of the First International
Workshop on Traffic Monitoring and Analysis, ser. TMA ’09. Berlin,
Heidelberg:
Springer-Verlag, 2009, pp. 26–33. [Online]. Available:
http://dx.doi.org/10.1007/978-3-642-01645-5 4
[83] G. Androulidakis, V. Chatzigiannakis, S. Papavassiliou, M. Grammatikou,
and V. Maglaris, “Understanding and evaluating the impact of
sampling on anomaly detection techniques,” in Proceedings of the
2006 IEEE Conference on Military Communications, ser. MILCOM’06.
Piscataway, NJ, USA: IEEE Press, 2006, pp. 349–355. [Online]. Available:
http://dl.acm.org/citation.cfm?id=1896579.1896633
[84] D. Brauckhoff, B. Tellenbach, A. Wagner, M. May, and A. Lakhina,
“Impact of packet sampling on anomaly detection metrics,” in Proceedings
of the 6th ACM SIGCOMM Conference on Internet Measurement, ser. IMC
’06. New York, NY, USA: ACM, 2006, pp. 159–164. [Online]. Available:
http://doi.acm.org/10.1145/1177080.1177101
BIBLIOGRAPHY
155
[85] A. Davy, D. Botvich, and B. Jennings, “On the use of accounting data for
qos-aware ip network planning,” in Proceedings of the 20th International
Teletraffic Conference on Managing Traffic Performance in Converged
Networks, ser. ITC20’07. Berlin, Heidelberg: Springer-Verlag, 2007, pp.
348–360. [Online]. Available: http://dl.acm.org/citation.cfm?id=1769187.
1769229
[86] Y.-s. Lim, H.-c. Kim, J. Jeong, C.-k. Kim, T. T. Kwon, and Y. Choi,
“Internet traffic classification demystified:
On the sources of the
discriminative power,” in Proceedings of the 6th International COnference,
ser. Co-NEXT ’10. New York, NY, USA: ACM, 2010, pp. 9:1–9:12.
[Online]. Available: http://doi.acm.org/10.1145/1921168.1921180
[87] Is See5/C5.0 Better Than C4.5?, [Online]. Available: http://rulequest.com/
see5-comparison.html, as of October 31, 2014.
[88] J. Li, S. Zhang, C. Li, and J. Yan, “Composite lightweight traffic
classification system for network management,” Int. J. Netw. Manag.,
vol. 20, no. 2, pp. 85–105, Mar. 2010. [Online]. Available: http:
//dx.doi.org/DOI:10.1002/nem.735
[89] A. Dainotti, A. Pescapé, and C. Sansone, “Early classification of
network traffic through multi-classification,” in Proceedings of the Third
International Conference on Traffic Monitoring and Analysis, ser. TMA’11.
Berlin, Heidelberg: Springer-Verlag, 2011, pp. 122–135. [Online]. Available:
http://dl.acm.org/citation.cfm?id=1986282.1986297
[90] S. Lee, H. Kim, D. Barman, S. Lee, C.-k. Kim, T. Kwon, and Y. Choi,
“Netramark: A network traffic classification benchmark,” SIGCOMM
Comput. Commun. Rev., vol. 41, no. 1, pp. 22–30, Jan. 2011. [Online].
Available: http://doi.acm.org/10.1145/1925861.1925865
[91] G. Hulten, L. Spencer, and P. Domingos, “Mining time-changing data
streams,” in Proceedings of the Seventh ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining, ser. KDD ’01.
New York, NY, USA: ACM, 2001, pp. 97–106. [Online]. Available:
http://doi.acm.org/10.1145/502512.502529
[92] W. Hoeffding, “Probability inequalities for sums of bounded random variables,” Journal of the American statistical association, vol. 58, no. 301, pp.
13–30, 1963.
156
BIBLIOGRAPHY
[93] A. Bifet and R. Gavaldà, “Adaptive learning from evolving data streams,”
in Proceedings of the 8th International Symposium on Intelligent Data
Analysis: Advances in Intelligent Data Analysis VIII, ser. IDA ’09.
Berlin, Heidelberg: Springer-Verlag, 2009, pp. 249–260. [Online]. Available:
http://dx.doi.org/10.1007/978-3-642-03915-7 22
[94] A. Bifet and R. Gavaldà, “Learning from time-changing data with adaptive
windowing,” in Siam International Data Mining Conference, 2007, pp. 443–
448.
[95] NBAR2 or Next Generation NBAR - Cisco, [Online]. Available:
http://www.cisco.com/en/US/prod/collateral/iosswrel/ps6537/ps6558/
ps6616/qa c67-697963.html, as of October 31, 2014.
[96] MOA: Massive Online Analysis, [Online]. Available:
waikato.ac.nz/, as of October 31, 2014.
http://moa.cms.
[97] A. Bifet and R. Kirkby, Data stream mining a practical approach. Citeseer,
2009.
[98] J. Gama, “A survey on learning from data streams: current and future
trends,” Progress in Artificial Intelligence, vol. 1, no. 1, pp. 45–55, 2012.
[99] J. a. Gama, R. Sebastião, and P. P. Rodrigues, “Issues in evaluation of
stream learning algorithms,” in Proceedings of the 15th ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining, ser.
KDD ’09. New York, NY, USA: ACM, 2009, pp. 329–338. [Online].
Available: http://doi.acm.org/10.1145/1557019.1557060
[100] Bujlow, T. et al., “Comparison of Deep Packet Inspection (DPI)
Tools for Traffic Classification,” UPC BarcelonaTech, Tech. Rep., 2013,
[Online]. Available: https://www.ac.upc.edu/app/research-reports/html/
research center index-CBA-2013,en.html, as of October 31, 2014.
[101] “Volunteer-Based System for Research on the Internet,” [Online]. Available:
http://vbsi.sourceforge.net/, as of October 31, 2014.
[102] J. Ma, K. Levchenko, C. Kreibich, S. Savage, and G. M. Voelker,
“Unexpected means of protocol inference,” in Proceedings of the 6th
ACM SIGCOMM Conference on Internet Measurement, ser. IMC ’06.
BIBLIOGRAPHY
157
New York, NY, USA: ACM, 2006, pp. 313–326. [Online]. Available:
http://doi.acm.org/10.1145/1177080.1177123
[103] C. Shen and L. Huang, “On detection accuracy of l7-filter and
opendpi,” in Proceedings of the 2012 Third International Conference on
Networking and Distributed Computing, ser. ICNDC ’12. Washington,
DC, USA: IEEE Computer Society, 2012, pp. 119–123. [Online]. Available:
http://dx.doi.org/10.1109/ICNDC.2012.36
[104] S. Alcock and R. Nelson, “Measuring the accuracy of open-source payloadbased traffic classifiers using popular internet applications,” in Local Computer Networks Workshops (LCN Workshops), 2013 IEEE 38th Conference
on. IEEE, 2013, pp. 956–963.
[105] M. Dusi, F. Gringoli, and L. Salgarelli, “Quantifying the accuracy
of the ground truth associated with internet traffic traces,” Comput.
Netw., vol. 55, no. 5, pp. 1158–1167, Apr. 2011. [Online]. Available:
http://dx.doi.org/10.1016/j.comnet.2010.11.006
[106] F. Gringoli, L. Salgarelli, M. Dusi, N. Cascarano, F. Risso, and k. c. claffy,
“Gt: Picking up the truth from the ground for internet traffic,” SIGCOMM
Comput. Commun. Rev., vol. 39, no. 5, pp. 12–18, Oct. 2009. [Online].
Available: http://doi.acm.org/10.1145/1629607.1629610
[107] R. Braden, “RFC 1122 - Requirements for Internet Hosts – Communication
Layers,” Internet Engineering Task Force, October 1989, [Online]. Available:
http://tools.ietf.org/html/rfc1122, as of October 31, 2014.
[108] V. Paxson, M. Allman, J. Chu, M. Sargent, “RFC 6298 - Computing TCP’s
Retransmission Timer,” Internet Engineering Task Force, June 2011, [Online]. Available: http://tools.ietf.org/html/rfc6298, as of October 31, 2014.
[109] W. John, S. Tafvelin, and T. Olovsson, “Trends and differences in
connection-behavior within classes of internet backbone traffic,” in
Proceedings of the 9th International Conference on Passive and Active
Network Measurement, ser. PAM’08. Berlin, Heidelberg: Springer-Verlag,
2008, pp. 192–201. [Online]. Available: http://dl.acm.org/citation.cfm?id=
1791949.1791976
[110] K. C. Claffy, H. W. Braun, and G. C. Polyzos, “A parameterizable
methodology for internet traffic flow profiling,” IEEE J.Sel. A. Commun.,
158
BIBLIOGRAPHY
vol. 13, no. 8, pp. 1481–1494, Sep. 2006. [Online]. Available:
//dx.doi.org/10.1109/49.464717
http:
[111] A. Dainotti, W. Donato, and A. Pescapé, “Tie: A community-oriented
traffic classification platform,” in Proceedings of the First International
Workshop on Traffic Monitoring and Analysis, ser. TMA ’09. Berlin,
Heidelberg:
Springer-Verlag, 2009, pp. 64–74. [Online]. Available:
http://dx.doi.org/10.1007/978-3-642-01645-5 8
[112] J. H. Friedman, J. L. Bentley, and R. A. Finkel, “An algorithm for
finding best matches in logarithmic expected time,” ACM Trans. Math.
Softw., vol. 3, no. 3, pp. 209–226, Sep. 1977. [Online]. Available:
http://doi.acm.org/10.1145/355744.355745
[113] J. L. Bentley, “K-d trees for semidynamic point sets,” in Proceedings
of the Sixth Annual Symposium on Computational Geometry, ser. SCG
’90. New York, NY, USA: ACM, 1990, pp. 187–197. [Online]. Available:
http://doi.acm.org/10.1145/98524.98564
Appendix A
159
160
8.2
UPC dataset
Next, we present the relation of collaborations related to the UPC dataset.
Supervisor
Institution
Date
Domenico Vitali
Jan 2011
Evaluation of traffic classification techniques.
Sanping Li
-
Feb 2011
My current research is mainly about online network traffic classification
based on multi-core architecture.
Qian Yaguan
Wu Chunming
Apr 2011
Our purpose is to analyze the details of mixed Internet traffic, and decide
proper policy for network recources distribution.
Yulios Zabala
Lee Luan Ling
Aug 2011
Massimiliano
Natale
Domenico Vitali
The purpose of our research is to develop a new technique for classification of network traffic based on multifractal theory and machine learning
algorithms, along with techniques to make the system usable in real time.
Traffic classification based on Netflow data.
Elie Bursztein
-
Jesus Diaz
Verdejo
-
Ning Gao
Quin Lv
Wesley Melo
Stenio Fernandes
Adriel Cheng
-
Corey Hart
-
Rajesh NP
-
Universita
degli
Studi di Roma La
Sapienza
(Rome,
Italy)
University of Massachusetts
Lowell
(Lowell, USA)
CS College of Zhejiang
University
(Hangzhou, China)
State University of
Campinas - Unicamp
(So Paulo, Brazil)
Universita
degli
Studi di Roma La
Sapienza
(Rome,
Italy)
Stanford University
(Stanford, USA)
Universidad
de
Granada (Granada,
Spain)
University of Colorado
Boulder
(Boulder, USA)
GPRT - Networking
and Telecommunications Research Group
(Recife, Brazil)
Department of Defence
(Edinburgh,
Australia)
Lockheed
Martin
(King of Prussia, PA,
USA)
Cisco (Bangalore, India)
Stanford University
(Stanford, USA)
Cisco (Bangalore, India)
Raja Rajendran
Andrew Ng
Indranil Adak
Raja Rajendran
Jan 2012
Feb 2012
Research Purpose
Feb 2013
Using the dataset to evaluate the effectiveness of our algorithm that reconstruct the statistical distribution of one classifier into another.
Validation of traffic classification techniques.
Feb 2013
I’m now doing a project to classify network flows into different applications.
Jul 2013
On my previous work, I used synthetic traces and I would like to evaluate
it using your real traces.
Sep 2013
Investigation of machine-learning/data-mining techniques for classification
of flows data.
Oct 2013
Spiking neural nework based intrusion detection.
Dec 2013
Evaluating multiple machine learning algorithms currently available for
Network analysis.
I would like to experiment with some of the machine learning algorithms
from my course using the netflow dataset.
We are currently researching on applying machine learning to predict the
traffic flow and recommending network/traffic growth for service providers.
Dec 2013
Dec 2013
APPENDIX A
Name
Giantonio
Chiarelli
PAM dataset
Next, we present the relation of collaborations related to the PAM dataset.
Name
Supervisor
Institution
Date
Research Purpose
Said Sedikki
Ye-Qiong Song
Mar 2014
Testing flow classification program that will be used in the controller for
an OpenFlow Network.
Oliver Gasser
Georg Carle
Mar 2014
We are doing research in DOS detection. Your dataset would be used for
background traffic purposes.
Viktor Minorov
Pavel Celada
Mar 2014
Data will be used in work about measuring the accuracy of traffic classifiers.
Yiyang Shao
Jun Li
University of Lorraine
(Villers-lesNancy, France)
Technische
Universitat
Munchen
(Munchen,
Germany)
Masaryk
University (Brno,
Czech
Republic)
Tsinghua University
(Beijing, China)
Apr 2014
Yinsen Miao
Farinaz
Koushanfar
Christof Fetzer
The ground-truth of traffic data is a fundamentally important issue in the
research of this area. So I hope I could have a copy of PAM dataset under
CBA privacy policies.
Apply this dataset to test our traffic classification method’s accuracy.
Le Quoc Do
Zuleika
Nascimento
Djamel Salok
Garrett Cullity
Adriel Cheng
Zeynab Sabahi
Joseph Kampeas
Ahmad
Nickabadi
Omer Gurewitz
Hossein Doroud
Andres Marin
Alioune BA
Cedric Baudoin
Jan-Erik Stange
-
Rice
University
(Houston, USA)
Technische Universitat Dresden (Dresden, Germany)
Federal University of
Pernambuco (Recife,
Brazil)
University of Adelaide (Adelaide, Australia)
University of Tehran
(Tehran, Iran)
Ben Gurion University of the Negev
(Beer Sheva, Israel)
Univesidad Carlos III
(Madrid, Spain)
Thales Alenia Space
(Toulouse, France)
University of Applied
Science
Postdam
(Postdam, Germany)
Apr 2014
May 2014
Optimizing Support Vector Machine (SVM) to improve classification accuracy in analyzing network data.
May 2014
I’m currently working on network traffic classification using soft computing
techniques and I’d like to test my model on different datasets.
May 2014
We aim to apply machine learning techniques to network flow data (netflow/IPFIX) for traffic classification, device characterisation, and hopefully, identification of malicious traffic.
I have implemented a new hybrid DPI tool based on data fusion techniques.
I need to compare the results of my tool with the existing DPI tools.
Traffic classification and identification using machine learning technique.
Jun 2014
Jun 2014
Jul 2014
Jul 2014
Jul 2014
8.3. PAM DATASET
8.3
Evaluate performance of a new network traffic classifier that uses Machine
Learning approach.
Use of DPI for dynamic QoS architecture. The aim is to validate the feasibility of DPI approach for crosslayer QoS optimization in satcom network.
The dataset will be used for visualization research in the EU-project
SaSER. More specifically our research team will use this dataset as a basis
for developing NetFlow visualization that assist in the process of detecting
brute-force and amplification attacks.
161
162
APPENDIX A
Appendix B
8.4
8.4.1
Publications
Journals
• V. Carela-Español, P. Barlet-Ros, A. Bifet and K. Fukuda. “A streaming
flow-based technique for traffic classification applied to 12+1 years of Internet
traffic”. Journal of Telecommunications Systems, 2014.(Under Review)
• T. Bujlow, V. Carela-Español and P. Barlet-Ros. “Independent Comparison of Popular DPI Tools for Trac Classification”. Computer Networks,
2014 (Under Review)
• V. Carela-Español, P. Barlet-Ros, O. Mula-Valls and J. Solé-Pareta. “An
Autonomic Traffic Classification System for Network Operation and Management”. Journal of Network and Systems Management, October 2013.
• V. Carela-Español, P. Barlet-Ros, A. Cabellos-Aparicio, and J. Solé-Pareta.
“Analysis of the impact of sampling on NetFlow traffic classification”. Computer Networks 55 (2011), pp. 1083-1099.
8.4.2
Conferences
• V. Carela-Español, T. Bujlow, and P. Barlet-Ros. “Is Our Ground-Truth
for Traffic Classification Reliable? ”. In Proc. of the Passive and Active
Measurements Conference (PAM’14), Los Angeles, CA, USA, March 2014.
• J. Molina, V. Carela-Español, R. Hoffmann, K. Degner and P. BarletRos. “Empirical analysis of traffic to establish a profiled flow termination
timeout”. In Proc. of 9th IEEE Intl. Wireless Communications and Mobile
Computing Conference (IWCMC-TRAC), Cagliari, Italy, July 2013.
163
164
APPENDIX B
• V. Carela-Español, P. Barlet-Ros, M. Solé-Simó, A. Dainotti, W. de Donato and A. Pesacapé. “K-dimensional trees for continuous traffic classification”. In Proc. of Second International Workshop on Traffic Monitoring
and Analysis. Zurich, Switzerland, April 2010.
• P. Barlet-Ros, V. Carela-Español, E. Codina and J. Solé-Pareta. “Identification of Network Applications based on Machine Learning Techniques”.
In Proc. of TERENA Networking Conference. Brugge, Belgium, May 2008.
8.4.3
Technical Reports
• T. Bujlow, V. Carela-Español and P. Barlet-Ros. “Extended Independent
Comparison of Popular Deep Packet Inspection (DPI) Tools for Traffic Classification”. Technical Report UPC-DAC-RR-CBA-2014-1, Jan. 2014.
• T. Bujlow, V. Carela-Español and P. Barlet-Ros. “Comparison of Deep
Packet Inspection (DPI) tools for traffic classification”. Technical Report
UPC-DAC-RR-CBA-2013-3, June 2013.
• V. Carela-Español, P. Barlet-Ros and J. Solé-Pareta. “Traffic classification with Sampled NetFlow ”. Technical Report UPC-DAC-RR-CBA-2009-6.
February 2009.
• V. Carela-Español, P. Barlet-Ros and J. Solé-Pareta. “Identification of
Network Applications based on Machine Learning Techniques”. Technical
Report UPC-DAC-RR-CBA-2009-2. February 2009.
8.4.4
Supervised Master Students
• Juan Molina Rodriguez : “Empirical analysis of traffic to establish a profiled
flow termination timeout”, 2013 in collaboration with ipoque.
8.4.5
Datasets
• UPC Dataset: NetFlow v5 dataset labeled by L7-filter
• PAM Dataset: full packet payload dataset labeled by VBS

Similar documents